UnifyBench

RankingKimi K3

Kimi K3

Compare models

Data updated 22 Oct 2026

81 published benchmark measures · 14 benchmark families contribute across 5 task areas. 4 capability estimates available in the full profile. See all results ↓ · Compare published benchmarks →

Model evidence summary

This profile combines published settings. It is not a runnable configuration or a leaderboard rank. Compare measured configurations →

Performance profile

Capabilities

Adjusted comparison score · 0–100
Kimi K3 capability radarComparison scores from 0 to 100. Unknown capabilities have no point; lines stop at gaps. Hollow points are preliminary. Exact values and support labels follow the chart.AgenticHard reasoningCodingHuman prefKnowledgeMultimodalLong context50100Kimi K3 · Agentic: 85.2 · SupportedKimi K3 · Hard reasoning: 54.3 · SupportedKimi K3 · Coding: 74.9 · PreliminaryKimi K3 · Multimodal: 76.4 · Supported

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

Agentic7 families · 1 with independent evidence · Supported85.2

7 core families; 48 direct opponents across 16 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: 75.0–87.4; 0/13 scenarios unsupported. Without one publisher: 78.9–89.0; 0/15 unsupported. Smoothing check: 77.2–89.1; 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.

Hard reasoning3 families · 2 with independent evidence · Supported54.3

3 core families; 63 direct opponents across 9 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: 41.5–72.2; 0/6 scenarios unsupported. Without one publisher: 42.9–58.1; 0/16 unsupported. Smoothing check: 52.7–55.4; 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.

Coding1 families · 1 with independent evidence · Preliminary74.9

1 core families; 27 direct opponents across 9 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; 6/6 scenarios unsupported. Without one publisher: No supported estimate; 19/19 unsupported. Smoothing check: No supported estimate; 3/3 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.

Human pref1 families · 1 with independent evidence · PreliminaryUnknown

1 core families; 76 direct opponents across 16 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; 1/1 scenarios unsupported. Without one publisher: No supported estimate; 1/1 unsupported. Smoothing check: No supported estimate; 3/3 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): 94.7 observed win share
    lmarena.ai · Source 1
KnowledgeNo comparable evidenceUnknown

0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

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.

    Multimodal2 families · 0 with independent evidence · Supported76.4

    2 core families; 4 direct opponents across 2 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: 75.4–77.8; 2/8 scenarios unsupported. Without one publisher: 75.1–79.1; 1/8 unsupported. Smoothing check: 70.8–79.1; 0/3 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.

    Long contextNo comparable evidenceUnknown

    0 core families; 0 direct opponents across 0 labs. No admitted core comparisons

    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.

      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.

      Reported effort · Max + 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.

      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
      Score contributions and missing evidence

      14 contributing families across 5 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
      1,000,000 tokens
      License
      —
      Model card
      https://huggingface.co/moonshotai/Kimi-K3
      Default Capability family coverage
      Documented-evidence family coverage/badge/kimi-k3.svg
      Benchmark scores & sources

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

      BenchmarkBucketScoreHarnessEvidenceSource-recorded date
      CyberGym · source release snapshot; version not specifiedSupporting evidence80%
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details.

      First-party reported result.

      zai-org/GLM-5.3 · Performance table: CyberGym · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      		lab self-report

      Narrow specialist biological, medical or security evaluation; preserved as raw evidence outside general capability scoring.

      		Reviewed 2026-10-18
      DAYJOB FinanceSupporting evidence3.8%DAYJOB Finance · max

      Effort: Not specified

      		official board

      Supporting evidence outside the reviewed capability core

      		2026-10-21
      DAYJOB HealthcareSupporting evidence1.2%DAYJOB Healthcare · max

      Effort: Not specified

      		official board

      Supporting evidence outside the reviewed capability core

      		2026-10-21
      DeepSearchQA (F1) · source release snapshot; version not specifiedAgentic95%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      First-party reported result.

      moonshotai/Kimi-K3 · Performance table: DeepSearchQA (F1) · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      contributes to capability      		lab self-reportReviewed 2026-10-18
      DeepSWE · source release snapshot; version not specifiedCoding67.5%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      First-party reported result.

      moonshotai/Kimi-K3 · Performance table: DeepSWE · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      contributes to capability      		lab self-reportReviewed 2026-10-18
      DeepSWE · source release snapshot; version not specifiedCoding67.5%
      Reported settings & source

      DeepSeek updated 0813/0731 release evaluation table; tool/harness settings per model card.

      First-party reported result.

      deepseek-ai/DeepSeek-V4-Pro-0813 · Performance table: DeepSWE · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      contributes to capability      		lab self-reportReviewed 2026-10-18
      DeepSWE (v1.1) · v1.1Coding67.5%
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details.

      First-party reported result.

      zai-org/GLM-5.3 · Performance table: DeepSWE (v1.1) · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      contributes to capability      		lab self-reportReviewed 2026-10-18
      DeepSWE v1.1Coding68.5%DeepSWE v1.1 reported

      Effort: Not specified

      contributes to capability      		official board2026-09-22
      DSBench-FullStack † · source release snapshot; version not specifiedSupporting evidence73.7%
      Reported settings & source

      DeepSeek updated 0813/0731 release evaluation table; tool/harness settings per model card.

      First-party reported result. † source footnote applies.

      deepseek-ai/DeepSeek-V4-Pro-0813 · Performance table: DSBench-FullStack † · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      		lab self-report

      Supporting evidence outside the reviewed capability core

      		Reviewed 2026-10-18
      DSBench-Hard † · source release snapshot; version not specifiedSupporting evidence63%
      Reported settings & source

      DeepSeek updated 0813/0731 release evaluation table; tool/harness settings per model card.

      First-party reported result. † source footnote applies.

      deepseek-ai/DeepSeek-V4-Pro-0813 · Performance table: DSBench-Hard † · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      		lab self-report

      Supporting evidence outside the reviewed capability core

      		Reviewed 2026-10-18
      ExploitBench · source release snapshot; version not specifiedSupporting evidence32.2%
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details.

      First-party reported result.

      zai-org/GLM-5.3 · Performance table: ExploitBench · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      		lab self-report

      Narrow specialist biological, medical or security evaluation; preserved as raw evidence outside general capability scoring.

      		Reviewed 2026-10-18
      ExploitGym (2h / 6h) · source release snapshot; version not specifiedSupporting evidence36 count
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details. 2h Single-run869tasks; API time normalized by model TPS plus overhead; reported number solved; domain whitelist.

      First-party reported result.

      zai-org/GLM-5.3 · Performance table: ExploitGym (2h / 6h) · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Narrow specialist biological, medical or security evaluation; preserved as raw evidence outside general capability scoring.

      		Reviewed 2026-10-18
      ExploitGym (2h / 6h) · source release snapshot; version not specifiedSupporting evidence70 count
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details. 6h Single-run869tasks; API time normalized by model TPS plus overhead; reported number solved; domain whitelist.

      First-party reported result.

      zai-org/GLM-5.3 · Performance table: ExploitGym (2h / 6h) · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Narrow specialist biological, medical or security evaluation; preserved as raw evidence outside general capability scoring.

      		Reviewed 2026-10-18
      Finance Agent v2 · source release snapshot; version not specifiedSupporting evidence54.4%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      Cited result; see benchmark-specific Evaluation Details.

      Comparison limit: Kimi K3 Evaluation Details cites Vals AI; retain as published context, not a new Moonshot comparison.

      moonshotai/Kimi-K3 · Performance table: Finance Agent v2 · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Kimi K3 Evaluation Details cites Vals AI; retain as published context, not a new Moonshot comparison.

      		Reviewed 2026-10-18
      FrontierMath ErdősSupporting evidence0%FrontierMath Erdős proof scorer

      Effort: Not specified

      		official board

      Supporting evidence outside the reviewed capability core

      		2026-08-28
      FrontierSWE · source release snapshot; version not specifiedSupporting evidence81.2%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      First-party reported result.

      Comparison limit: Historical provider claim; secondary and superseded for current FrontierSWE comparisons by the independent Proximal FrontierSWE v2 mean@5 board at https://www.frontierswe.com/. Version-unspecified claims are not v2 results. Excluded from weighted Overall.

      moonshotai/Kimi-K3 · Performance table: FrontierSWE · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Historical provider claim; secondary and superseded for current FrontierSWE comparisons by the independent Proximal FrontierSWE v2 mean@5 board at https://www.frontierswe.com/. Version-unspecified claims are not v2 results. Excluded from weighted Overall.

      		Reviewed 2026-10-18
      FrontierSWE v2Supporting evidence25.9%Proximal FrontierSWE v2 · Proximus · unspecified

      Effort: Not specified

      		official board

      Supporting evidence outside the reviewed capability core

      		2026-10-22
      GDPval-AAAgentic1533Artificial Analysis GDPval-AA

      Effort: Not specified

      contributes to capability      		official board2026-10-22
      GDPval-AA v2 · source release snapshot; version not specifiedAgentic1682
      Reported settings & source

      GLM-5.3 release table; benchmark-specific harness and reasoning configurations in Evaluation Details.

      Externally evaluated result, attributed in the source footnote.

      Comparison limit: The GLM card attributes this evaluation to Artificial Analysis; it is not a new Z.ai comparison.

      zai-org/GLM-5.3 · Performance table: GDPval-AA v2 · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      		lab self-report

      The GLM card attributes this evaluation to Artificial Analysis; it is not a new Z.ai comparison.

      		Reviewed 2026-10-18
      GDPval-AA v2 (Elo) · source release snapshot; version not specifiedAgentic1686
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      Cited result; see benchmark-specific Evaluation Details.

      Comparison limit: Kimi K3 Evaluation Details cites Artificial Analysis; retain as published context, not a new Moonshot comparison.

      moonshotai/Kimi-K3 · Performance table: GDPval-AA v2 (Elo) · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Kimi K3 Evaluation Details cites Artificial Analysis; retain as published context, not a new Moonshot comparison.

      		Reviewed 2026-10-18
      GPQA DiamondHard reasoning93.535%GPQA Diamond reported

      Effort: Not specified

      contributes to capability      		independent repro2026-10-20
      GPQA Diamond · source release snapshot; version not specifiedHard reasoning93.5%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      First-party reported result.

      moonshotai/Kimi-K3 · Performance table: GPQA Diamond · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      contributes to capability      		lab self-reportReviewed 2026-10-18
      Harvey Lab-AA · source release snapshot; version not specifiedSupporting evidence94.6%
      Reported settings & source

      Table headers: Kimi K3, GPT-5.6 Sol, Opus 4.8 and GLM-5.2 max; GPT-5.5 xhigh; Fable 5 max with fallbacks. Benchmark-specific footnotes override header settings; cited results are not new Moonshot evaluations. Benchmark-specific tools, harness and budget documented under Evaluation Details. Kimi temp1; single-step top_p.95,agentic top_p1; vision tools=Python,3runs except ZeroBench5runs.

      Cited result; see benchmark-specific Evaluation Details.

      Comparison limit: Kimi K3 Evaluation Details cites Artificial Analysis; retain as published context, not a new Moonshot comparison.

      moonshotai/Kimi-K3 · Performance table: Harvey Lab-AA · reviewed 2026-10-18

      		Published configuration

      Effort: Max

      		lab self-report

      Kimi K3 Evaluation Details cites Artificial Analysis; retain as published context, not a new Moonshot comparison.

      		Reviewed 2026-10-18
      HLE (wo / w tools) · source release snapshot; version not specifiedHard reasoning43.5%
      Reported settings & source

      DeepSeek updated 0813/0731 release evaluation table; tool/harness settings per model card. without tools

      First-party reported result.

      deepseek-ai/DeepSeek-V4-Pro-0813 · Performance table: HLE (wo / w tools) · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      contributes to capability      		lab self-reportReviewed 2026-10-18
      HLE (wo / w tools) · source release snapshot; version not specifiedHard reasoning56%
      Reported settings & source

      DeepSeek updated 0813/0731 release evaluation table; tool/harness settings per model card. with tools

      First-party reported result.

      deepseek-ai/DeepSeek-V4-Pro-0813 · Performance table: HLE (wo / w tools) · reviewed 2026-10-18

      		Published configuration

      Effort: Not specified

      contributes to capability      		lab self-reportReviewed 2026-10-18