Kimi-K2.6

Data updated 12 Sept 2026

66 published benchmark measures · 19 benchmark families contribute across 7 task areas. 5 capability estimates available in the full profile. See all results ↓ · Compare published benchmarks →

Model evidence summary

Choose a configuration

A model-wide score would mix effort settings. This page summarizes published evidence; only identified configurations receive leaderboard ranks.

Compare measured configurations →

The capability chart summarizes evidence across reported settings, not a runnable configuration or a leaderboard rank.

Performance profile

Capabilities

Adjusted comparison score · 0–100

Filled points are supported; hollow points are preliminary. Lines stop at unknown capabilities. Exact values and sources follow below.

Agentic4 families · 0 with independent evidence · Supported12.8

4 core families; 9 direct opponents across 7 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 9.0–15.6; 0/11 scenarios unsupported. Without one publisher: 11.1–95.8; 0/15 unsupported. Smoothing check: 9.4–20.7; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage; source sensitivity still requires inspection. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Browsecomp: 78.6 observed win share
MiniMax, NVIDIA · Source 1 Source 2
GDPval: 44.6 observed win share
MiniMax, NVIDIA · Source 1 Source 2
OSWorld: 20.0 observed win share
MiniMax · Source 1
MCP-Atlas: 28.6 observed win share
MiniMax · Source 1

Hard reasoning4 families · 0 with independent evidence · Supported38.5

4 core families; 8 direct opponents across 7 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 33.1–46.6; 0/7 scenarios unsupported. Without one publisher: 16.5–66.7; 1/16 unsupported. Smoothing check: 30.3–45.7; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Aime: 100.0 observed win share
Microsoft · Source 1
GPQA: 80.0 observed win share
Microsoft, NVIDIA · Source 1 Source 2
Humanity’s Last Exam: 100.0 observed win share
NVIDIA · Source 1 Source 2
Imo Answerbench: 100.0 observed win share
NVIDIA · Source 1 Source 2

Coding6 families · 0 with independent evidence · Supported11.1

6 core families; 12 direct opponents across 8 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: 5.2–38.5; 0/7 scenarios unsupported. Without one publisher: 8.1–34.3; 0/20 unsupported. Smoothing check: 8.5–17.9; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts improved on neutral predictions within evaluated coverage. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

SWE-bench Multilingual: 50.0 observed win share
Moonshot AI, NVIDIA · Source 1 Source 2
Terminal-Bench: 88.9 observed win share
MiniMax, Moonshot AI, NVIDIA · Source 1 Source 2 Source 3
LiveCodeBench: 100.0 observed win share
Microsoft, NVIDIA · Source 1 Source 2
SWE-bench Pro: 79.2 observed win share
Microsoft, MiniMax · Source 1 Source 2
SWE-bench Verified: 58.3 observed win share
Microsoft, MiniMax, NVIDIA · Source 1 Source 2 Source 3
Scicode: 100.0 observed win share
NVIDIA · Source 1

Human pref1 families · 1 with independent evidence · Supported33.7

1 core families; 114 direct opponents across 19 labs.

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 1/1 scenarios unsupported. Without one publisher: No supported estimate; 1/1 unsupported. Smoothing check: 25.0–41.5; 0/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: cross-family and cross-publisher predictions could not be evaluated from the available comparison network. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Human preference (Arena): 83.3 observed win share
lmarena.ai · Source 1

Knowledge1 families · 0 with independent evidence · PreliminaryUnknown

1 core families; 4 direct opponents across 4 labs. No connected comparison to the complete reference panel

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 4/4 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: almost no publisher-held-out comparisons remained estimable. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

MMLU-Pro: 62.5 observed win share
NVIDIA · Source 1 Source 2

Multimodal2 families · 0 with independent evidence · Preliminary12.9

2 core families; 6 direct opponents across 4 labs. Needs broader benchmark and opponent support

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 8/8 scenarios unsupported. Without one publisher: No supported estimate; 8/8 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: limited publisher-held-out coverage. Cross-source validity is unresolved. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

CharXiv: 50.0 observed win share
Moonshot AI · Source 1
Mmmu Pro: 30.0 observed win share
MiniMax, Moonshot AI · Source 1 Source 2

Long context1 families · 0 with independent evidence · PreliminaryUnknown

1 core families; 4 direct opponents across 4 labs. No connected comparison to the complete reference panel

Estimated against the same complete panel: gpt-5.6-sol claude-opus-4-8 gemini-3.1-pro . Indirect comparisons assume performance can be summarized by one strength within this capability.

Refit without one family: No supported estimate; 5/5 scenarios unsupported. Without one publisher: No supported estimate; 5/5 unsupported. Smoothing check: No supported estimate; 3/3 unsupported. These are sensitivity checks, not confidence intervals.

Audit: family and publisher holdouts underperformed neutral predictions. Treat this as an exploratory estimate. Validation audit.

Matched family win shares below describe source evidence; they are not averaged to produce the adjusted score.

Aa Lcr: 100.0 observed win share
NVIDIA · Source 1

Scores estimate outcomes against a shared reference panel; they are not accuracy percentages. Sparse or disconnected evidence cannot qualify an overall profile. Open a capability to inspect its evidence.

Compare 3 effort levels across 36 benchmark/harness combinations →

Reported effort · Reasoning + unspecified

Settings reported in this model's published benchmark results, including results outside the aggregate. Effort names are provider-specific. These are not API defaults or equal compute budgets.

Reasoning: 33 observations
Not specified: 53 observations

Mixed settings means multiple settings occur in the evidence. Best across efforts means the source selected its best reported result across settings; it does not mean Max. Unspecified settings stay unknown. This model-summary chart combines reported settings. The leaderboard keeps identified configurations separate and excludes unknown effort.

Inspect each result and its source ↓ · Download effort evidence

Compare capability profiles →

Score contributions and missing evidence

19 contributing families across 7 capabilities. Fixed reference panels do not change when the catalog expands.

Capability is fitted jointly across families. Capability estimates below describe different task areas; their weighted sum is not the Capability score.

Results without reviewed compatibility or a reference match remain in the raw evidence below. Coverage counts only contributing results.

Model information & shareable badge

Lab: Moonshot
Catalog status: api-listed
Availability: Documented provider API and open weights for self-hosting
Family: Kimi
Released: —
Context: 256,000 tokens
License: —
Model card: https://huggingface.co/moonshotai/Kimi-K2.6
Default Capability family coverage: /badge/kimi-k2.6.svg

Benchmark scores & sources

Original results, evaluation harnesses, and evidence behind this model.

Benchmark	Bucket	Score	Harness	Evidence	Source-recorded date
AA-LCR · source release snapshot; version not specified	Long context	70.2% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 70.2 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: AA-LCR · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
AIME 2026	Hard reasoning	96.4% Reported settings & source Kimi K2.6 card, AIME 2026. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Reasoning: 98304 generation tokens; HLE full set. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, AIME 2026, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
AIME 2026 · source release snapshot; version not specified	Hard reasoning	96.4% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
APEX-Agents	Agentic	27.9% Reported settings & source Kimi K2.6 card, APEX-Agents. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, APEX-Agents, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
Apex-Shortlist (no tools) · source release snapshot; version not specified	Supporting evidence	77.4% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 77.4 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Apex-Shortlist (no tools) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
Apex-Shortlist (with tools) · source release snapshot; version not specified	Supporting evidence	73.2% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 73.2 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Apex-Shortlist (with tools) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
BabyVision	Supporting evidence	39.8% Reported settings & source Kimi K2.6 card, BabyVision. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, BabyVision, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
BabyVision (w/ python)	Supporting evidence	68.5% Reported settings & source Kimi K2.6 card, BabyVision (w/ python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, BabyVision (w/ python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
BrowseComp	Agentic	83.2% Reported settings & source Kimi K2.6 card, BrowseComp. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, BrowseComp, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
BrowseComp · source release snapshot; version not specified	Agentic	83.2% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
BrowseComp · source release snapshot; version not specified	Agentic	61.3% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 61.3 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: BrowseComp · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
BrowseComp (Agent Swarm)	Agentic	86.3% Reported settings & source Kimi K2.6 card, BrowseComp (Agent Swarm). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Comparison limit: Agent swarm execution is not an individual model configuration. Kimi K2.6 official model card · Evaluation Results table, BrowseComp (Agent Swarm), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Agent swarm execution is not an individual model configuration.	Reviewed 2026-09-12
CharXiv (RQ)	Multimodal	80.4% Reported settings & source Kimi K2.6 card, CharXiv (RQ). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, CharXiv (RQ), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning contributes to capability	lab self-report	Reviewed 2026-09-12
CharXiv (RQ) (w/ python)	Supporting evidence	86.7% Reported settings & source Kimi K2.6 card, CharXiv (RQ) (w/ python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, CharXiv (RQ) (w/ python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
Claw Eval (pass@3)	Supporting evidence	80.9% Reported settings & source Kimi K2.6 card, Claw Eval (pass@3). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, Claw Eval (pass@3), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
Claw Eval (pass^3)	Supporting evidence	62.3% Reported settings & source Kimi K2.6 card, Claw Eval (pass^3). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, Claw Eval (pass^3), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
Claw-Eval · source release snapshot; version not specified	Supporting evidence	61.5% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
CritPt (no tools) · source release snapshot; version not specified	Supporting evidence	9.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 9.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: CritPt (no tools) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
DeepSearchQA (accuracy)	Supporting evidence	83% Reported settings & source Kimi K2.6 card, DeepSearchQA (accuracy). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, DeepSearchQA (accuracy), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
DeepSearchQA (f1-score)	Supporting evidence	92.5% Reported settings & source Kimi K2.6 card, DeepSearchQA (f1-score). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, DeepSearchQA (f1-score), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
GDPVal · source release snapshot; version not specified	Agentic	50.4% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 50.4 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: GDPVal · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GDPval rubrics · source release snapshot; version not specified	Agentic	65.1% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GPQA (no tools) · source release snapshot; version not specified	Hard reasoning	91% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 91.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: GPQA (no tools) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GPQA Diamond · source release snapshot; version not specified	Hard reasoning	90.5% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
GPQA-Diamond	Supporting evidence	90.5% Reported settings & source Kimi K2.6 card, GPQA-Diamond. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Reasoning: 98304 generation tokens; HLE full set. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, GPQA-Diamond, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
HLE (no tools) · source release snapshot; version not specified	Hard reasoning	34.8% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 34.8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: HLE (no tools) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
HLE (with tools) · source release snapshot; version not specified	Hard reasoning	54% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 54.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: HLE (with tools) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
HLE-Full	Hard reasoning	34.7% Reported settings & source Kimi K2.6 card, HLE-Full. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Reasoning: 98304 generation tokens; HLE full set. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, HLE-Full, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
HLE-Full (w/ tools)	Hard reasoning	54% Reported settings & source Kimi K2.6 card, HLE-Full (w/ tools). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, HLE-Full (w/ tools), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
HMMT 2026 (Feb)	Supporting evidence	92.7% Reported settings & source Kimi K2.6 card, HMMT 2026 (Feb). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Reasoning: 98304 generation tokens; HLE full set. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, HMMT 2026 (Feb), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
HMMT February 2026 · source release snapshot; version not specified	Supporting evidence	92.7% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
IFBench (prompt loose) · source release snapshot; version not specified	Supporting evidence	73.7% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 73.7 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: IFBench (prompt loose) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
IMO-AnswerBench	Supporting evidence	86% Reported settings & source Kimi K2.6 card, IMO-AnswerBench. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Reasoning: 98304 generation tokens; HLE full set. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, IMO-AnswerBench, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
IMOAnswerBench (no tools) · source release snapshot; version not specified	Hard reasoning	91.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 91.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: IMOAnswerBench (no tools) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
IMOAnswerBench (with tools) · source release snapshot; version not specified	Hard reasoning	93.71% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 93.71 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: IMOAnswerBench (with tools) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
IOI 2025 · source release snapshot; version not specified	Supporting evidence	585 points Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 585.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: IOI 2025 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
LiveCodeBench (v6)	Coding	89.6% Reported settings & source Kimi K2.6 card, LiveCodeBench (v6). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, LiveCodeBench (v6), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
LiveCodeBench (v6) · source release snapshot; version not specified	Coding	90.2% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 90.2 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: LiveCodeBench (v6) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
LiveCodeBench v6 · source release snapshot; version not specified	Coding	89.6% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
LMArena Text Arena	Human pref	1461	LMArena Text Effort: Not specified contributes to capability	official board	2026-09-11
MathVision	Supporting evidence	87.4% Reported settings & source Kimi K2.6 card, MathVision. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, MathVision, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
MathVision (w/ python)	Supporting evidence	93.2% Reported settings & source Kimi K2.6 card, MathVision (w/ python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, MathVision (w/ python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
MCPAtlas · source release snapshot; version not specified	Agentic	66.6% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MCPMark	Supporting evidence	55.9% Reported settings & source Kimi K2.6 card, MCPMark. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, MCPMark, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
MMLU-Pro · source release snapshot; version not specified	Knowledge	88.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 88.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: MMLU-Pro · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MMLU-ProX (avg en/de/fr/es/it/ja/zh/hi/pt/ko) · source release snapshot; version not specified	Knowledge	85% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 85.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: MMLU-ProX (avg en/de/fr/es/it/ja/zh/hi/pt/ko) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MMMU-Pro	Multimodal	79.4% Reported settings & source Kimi K2.6 card, MMMU-Pro. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, MMMU-Pro, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning contributes to capability	lab self-report	Reviewed 2026-09-12
MMMU-Pro · source release snapshot; version not specified	Multimodal	79.4% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
MMMU-Pro (w/ python)	Supporting evidence	80.1% Reported settings & source Kimi K2.6 card, MMMU-Pro (w/ python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, MMMU-Pro (w/ python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
Multi-Challenge · source release snapshot; version not specified	Supporting evidence	63.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 63.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Multi-Challenge · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
NL2Repo · source release snapshot; version not specified	Supporting evidence	42.8% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
OJBench (python)	Supporting evidence	60.6% Reported settings & source Kimi K2.6 card, OJBench (python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, OJBench (python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
OmniScience Accuracy · source release snapshot; version not specified	Supporting evidence	35.5% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 35.5 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: OmniScience Accuracy · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
OSWorld Verified · source release snapshot; version not specified	Agentic	73.1% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
OSWorld-Verified	Agentic	73.1% Reported settings & source Kimi K2.6 card, OSWorld-Verified. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, OSWorld-Verified, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
PinchBench · source release snapshot; version not specified	Supporting evidence	90.2% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 90.2 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: PinchBench · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
ProfBench (Search) · source release snapshot; version not specified	Supporting evidence	56% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 56.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: ProfBench (Search) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
SciCode	Coding	52.2% Reported settings & source Kimi K2.6 card, SciCode. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, SciCode, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
SciCode (subtask) · source release snapshot; version not specified	Coding	52% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 52.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: SciCode (subtask) · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SpreadsheetBench v1 · source release snapshot; version not specified	Supporting evidence	84.5% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
SVG-Bench · source release snapshot; version not specified	Supporting evidence	60% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
SWE-Bench Multilingual	Coding	76.7% Reported settings & source Kimi K2.6 card, SWE-Bench Multilingual. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, SWE-Bench Multilingual, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning contributes to capability	lab self-report	Reviewed 2026-09-12
SWE-Bench Multilingual · source release snapshot; version not specified	Coding	77.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 77.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: SWE-Bench Multilingual · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-Bench Pro	Coding	58.6% Reported settings & source Kimi K2.6 card, SWE-Bench Pro. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, SWE-Bench Pro, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
SWE-bench Pro · source release snapshot; version not specified	Coding	58.6% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-bench Pro · source release snapshot; version not specified	Coding	58.6% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-Bench Verified	Coding	80.2% Reported settings & source Kimi K2.6 card, SWE-Bench Verified. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, SWE-Bench Verified, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
SWE-bench Verified · source release snapshot; version not specified	Coding	80.2% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-bench Verified · source release snapshot; version not specified	Coding	80.2% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
SWE-Bench Verified · source release snapshot; version not specified	Coding	75.7% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 75.7 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: SWE-Bench Verified · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
TauBench V3 Airline · source release snapshot; version not specified	Supporting evidence	85.8% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 85.8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Airline · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
TauBench V3 Average · source release snapshot; version not specified	Supporting evidence	72.4% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 72.4 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Average · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
TauBench V3 Banking · source release snapshot; version not specified	Supporting evidence	23.1% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 23.1 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Banking · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
TauBench V3 Retail · source release snapshot; version not specified	Supporting evidence	82.9% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 82.9 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Retail · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
TauBench V3 Telecom · source release snapshot; version not specified	Supporting evidence	97.8% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 97.8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Telecom · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
Terminal Bench 2.1 · 2.1	Coding	67.2% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 67.2 The linked public reproduction recipe names Terminal Bench 2.0, while this card labels 2.1. Recipe settings are not transferred across that unresolved version mismatch. nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: Terminal Bench 2.1 · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
Terminal-Bench 2.0 · 2.0	Coding	66.7% Reported settings & source MAI avg 4 runs, temperature 1, top-p .97; simple ReAct bash/string-replace; Terminal-Bench ignores predefined timeouts. Comparator configurations from cited reports. First-party reported result; comparator results retain the source evaluation setup. MAI-Thinking-1 technical report · Table 11, PDF page 53 · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Microsoft report Table 11 and section 4.1: MAI Terminal-Bench removes timeouts and uses a minimal ReAct harness; comparator values are cited from other model releases. A common evaluation protocol is not established.	Reviewed 2026-09-06
Terminal-Bench 2.0 (Terminus-2)	Coding	66.7% Reported settings & source Kimi K2.6 card, Terminal-Bench 2.0 (Terminus-2). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Coding: average of ten runs; Terminal uses Terminus2 preserve-thinking; SWE uses in-house SWE-agent. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, Terminal-Bench 2.0 (Terminus-2), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning contributes to capability	lab self-report	Reviewed 2026-09-12
Terminal-Bench 2.1 · 2.1	Coding	53.9% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified contributes to capability	lab self-report	Reviewed 2026-09-06
Toolathlon	Agentic	50% Reported settings & source Kimi K2.6 card, Toolathlon. Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, Toolathlon, column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report No matched opponent in this evaluation unit	Reviewed 2026-09-12
V* (w/ python)	Supporting evidence	96.9% Reported settings & source Kimi K2.6 card, V* (w/ python). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Vision: 98304 tokens, average of three runs; Python condition uses 65536 tokens per step, at most50steps. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, V* (w/ python), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
Vals.ai Financial Agent 1.1 with web search · 1.1	Supporting evidence	58.8% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 58.8 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: with web search · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
Vals.ai Financial Agent 1.1 without web search · 1.1	Supporting evidence	54% Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 54.0 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: without web search · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
VIBE-V2 · source release snapshot; version not specified	Supporting evidence	46% Reported settings & source MiniMax launch figure methodology. SWE internal ClaudeCode 4 runs; Terminal 8CPU16GB2h128k Terminus2; Paper/PostTrain Ralph-loop12h; GDPval internal rubric grader; BrowseComp discard64k; OSWorld361tasks. Comparator provider/leaderboard scores where footnoted. First-party reported result; comparator results retain the source evaluation setup. MiniMax M3 model card benchmark figure · figures/benchmark.jpeg · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
WideSearch (item-f1)	Supporting evidence	80.8% Reported settings & source Kimi K2.6 card, WideSearch (item-f1). Thinking mode; temperature 1.0, top-p 1.0, context 262144. Comparator settings: GPT-5.4 xhigh, Opus4.6 max, Gemini3.1Pro high. See benchmark-specific footnotes. Only own K2.6 and starred re-evaluations can form new comparisons. Source SHA256 95db3be1d0473e482c7aa901f237ad7af8bb7929625a777b27428081bb8ea9f7. Own measurement or explicitly starred same-condition re-evaluation. Assess benchmark-specific protocol before admission. Kimi K2.6 official model card · Evaluation Results table, WideSearch (item-f1), column 2 · reviewed 2026-09-12	Published configuration Effort: Reasoning	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-12
WMT24++ (en→xx) · source release snapshot; version not specified	Supporting evidence	84.5 score (source scale) Reported settings & source NVIDIA evaluation harness/settings per benchmark in model card; comparator scores are NVIDIA-reported under that protocol. First-party reported result. Original cell: 84.5 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 · Performance table: WMT24++ (en→xx) · reviewed 2026-09-06	Published configuration Effort: Not specified	lab self-report Supporting evidence outside the reviewed capability core	Reviewed 2026-09-06
ARC-AGI-2	Hard reasoning	—	—	—	—
DeepSWE v1.1	Agentic	—	—	—	—
GDPval-AA	Agentic	—	—	—	—
GPQA Diamond	Hard reasoning	—	—	—	—
Humanity's Last Exam	Hard reasoning	—	—	—	—
LiveCodeBench	Coding	—	—	—	—
MMLU-Pro	Knowledge	—	—	—	—
OSWorld-Verified	Agentic	—	—	—	—
SWE-bench Pro	Agentic	—	—	—	—
SWE-bench Verified	Agentic	—	—	—	—
Terminal-Bench 2.1	Agentic	—	—	—	—