Models · Moonshot AI (Kimi) · Out sinceReleased 16 Jul 2026
Kimi K3#8 for writing.#8 for writing, best at max effort.
Available on OpenRouterOn OpenRouterAvailable nowGenerally availableCan run on your own serversOpen weights
Kimi K3 is made by Moonshot AI (Kimi). Among the models we track it ranks #8 for writing, #9 for coding, #12 for research and analysis. It's expensive to use. Its makers have published it, so you can run it on your own servers.
2.8T total / 104B active MoE (KDA + gated MLA, AttnRes), MXFP4 QAT, 1M ctx. Kimi K3 License (open weights released 2026-07-27). Thinking always on; reasoning_effort low/high/max (default max); at launch only max was available. Cache-hit input $0.30; cache write $3 (5 min) / $6 (1 h). Blog dated 2026-07-17 (UTC+8); OpenRouter listing 2026-07-16.
- 61.5 / 100 · #8
- 72.5 / 100 · #12
- 67.0 / 100 · #9
- Expensive$3 / $15
- $6
- 1M
- —
- 83 (62 independent62 indep.)
- 16 Jul 2026
- Moonshot AI (Kimi)
- text, image, video
Scores are out of 100 for each area, compared with every model we track; “#” is its rank. Context is how much text it can read at once (1M is roughly 700,000 words). Blended price mixes the cost of what you send and what it writes back.
LowHighMax
Highlighted: where it did best for writing (Maximum thinking).
Highlighted: dominant setting in its writing composite.
OpenRouter is a service that gives access to many AI models in one place. This model is listed there as moonshotai/kimi-k3.
Route it as moonshotai/kimi-k3 at $0.71 in / $10 out per 1M tokens, 1M context. Listed since 16 Jul 2026.
frequency_penaltyinclude_reasoninglogit_biaslogprobsmax_tokensmin_ppresence_penaltyreasoningreasoning_effortrepetition_penaltyresponse_formatseedstopstructured_outputstemperaturetool_choicetoolstop_ktop_logprobstop_p
Run it yourself
Self-hosting
On your own servers,Licence, sizeunder your own control.and hardware.
- Yes
- 2800B / 104B
- No
Needs several AI servers.
> 1.1 TB at 8-bit (multi-node). Weights ≈ 2940 GB at 8-bit, 1540 GB at 4-bit (+10–30% for KV cache). MoE: 104B active per token.
Licence conditions: Keep copyright/permission notice. If you (with affiliates) run a 'Model as a Service' business with >US$20M revenue over 12 months you need a separate agreement with Moonshot before any commercial use. Commercial products with >100M MAU or >US$20M monthly revenue must prominently display 'Kimi K3'. Neither condition applies to internal use (outputs/capabilities not made available to third parties).
Restrictions: Keep copyright/permission notice. If you (with affiliates) run a 'Model as a Service' business with >US$20M revenue over 12 months you need a separate agreement with Moonshot before any commercial use. Commercial products with >100M MAU or >US$20M monthly revenue must prominently display 'Kimi K3'. Neither condition applies to internal use (outputs/capabilities not made available to third parties).
Languages: Languages: Not stated in model card.
Training it further: Fine-tuning: No official fine-tuning recipe in the model card; license explicitly permits fine-tuning.
Quantisations: mxfp4 (native QAT weights, MXFP8 activations), nvfp4 (NVIDIA: nvidia/Kimi-K3-NVFP4). Engines: vLLM, SGLang, TokenSpeed.
Download (Hugging Face)Weights on Hugging Face ↗huggingface.co ↗huggingface.co ↗
Thinking level
Reasoning effort
How long should Kimi K3Where Kimi K3think?peaks.
Each chart is one test. Left to right, the model thinks longer; higher is a better result. The yellow ring marks where it did best. Thinking level is how long the AI thinks before it answers. Higher is slower and costs more, and it isn't always better.
Score per reasoning setting on each benchmark with ≥2 settings; ring = peak. Canonical results (independent over vendor).
GPQA Diamond
Best at MaxBest at Max
Every test result
Every result
All the numbers,with where they came from.
Every result we've found for this model, with who measured it and a link to where we read it.
All raw rows (vendor and independent kept separate), with setting label, source, date, harness and notes.
| NotesNotes | ||||||
|---|---|---|---|---|---|---|
| AA Analyst Agent | 38.8% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run; scores move in 1.25-pt steps (small task set) |
| AA-Briefcase v1.1 | 1501 | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | |
| AA-LCR | 88.7% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-LCR accuracy, AA-run (long multi-document reasoning) |
| AA-Omniscience Accuracy | 47.6% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | |
| AA-Omniscience Hallucination Rate | 53.2% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | share of non-correct answers that were wrong instead of abstaining; = 1 - AA 'omniscienceNonHallucination' field (equals breakdown.hallucinationRate on the eval page) |
| AA-Omniscience Index | 19.7 | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-Omniscience Index (-100..100): correct minus incorrect, abstentions not penalised |
| Agents' Last Exam | 28.3% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | Cited from official leaderboard (2026-07-23) |
| Artificial Analysis Coding Agent Index | 51.9 | DefaultKimi K3 | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | Kimi Code CLI | agent Kimi Code CLI; components: DeepSWE v1.1 68.4, SWE-Atlas-QnA 66.1, Terminal-Bench v4 21.2; avg cost $5.05/task; avg wall time 61 min/task |
| Artificial Analysis Intelligence Index | 43.6 | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA slug kimi-k3; list price $3/15 per 1M in/out; cost to run AA Intelligence Index $2.00/task |
| Artificial Analysis output speed | 34 tok/s | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | tokens/sec (median output speed, first-party API); TTFT 4.1s; list price $3/15 per 1M in/out |
| AutomationBench | 30.8% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | 600-task public subset |
| BrowseComp | 91.2% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | Context compaction at 300K tokens; 90.4 with full 1M ctx and no context management |
| DeepSWE | 68.5% | Maxkimi-k3_max | Independent testIndependentDeepSWE (Datacurve) via Epoch AI ↗ | 1 Oct 2026 | mini-SWE-agent | read from Epoch AI benchmark_data.zip (deepswe_external.csv); original leaderboard https://deepswe.datacurve.ai/; harness mini-swe-agent; pass@4 89.4%; ±4.5; 4 runs; $4.65/task |
| DeepSWE | 67.5% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | DeepSWE v1.1; 67.3 with mini-SWE-agent harness |
| DeepSWE | 68.4% | DefaultKimi K3 | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | Kimi Code CLI | AA Coding Agent Index component; pass@1 avg of 3 attempts |
| Design Arena (all categories) | 1372 | Defaultkimi-k3 | Independent testIndependentDesign Arena ↗ | 1 Oct 2026 | — | Bradley-Terry Elo; ±4.3 SE; 7614 battles; win rate 63.4% |
| Design Arena (fullstack) | 1314 | Defaultkimi-k3 | Independent testIndependentDesign Arena ↗ | 1 Oct 2026 | — | Bradley-Terry Elo; ±10.3 SE; 1311 battles; win rate 61.9% |
| Epoch Capabilities Index | 157.6 | Default | Independent testIndependentEpoch AI ↗ | 1 Oct 2026 | — | Epoch Capabilities Index; 90% CI 154.9-160.4; best of listed model versions |
| EQ-Bench Creative Writing v3 (Elo) | 2082 | Defaultkimi-k3 | Independent testIndependentEQ-Bench ↗ | 1 Oct 2026 | — | leaderboard rank 5; rubric score 16.85/20; slop 9.70; avg length 5488 chars; Elo judged by Claude Sonnet 4.6; effort not stated by EQ-Bench |
| FrontierCode | 44.2% | Defaultkimi-k3_unknown | Independent testIndependentFrontierCode (Cognition) via Epoch AI ↗ | 1 Oct 2026 | mini-SWE-agent | read from Epoch AI benchmark_data.zip (frontiercode_external.csv); original leaderboard https://cognition.com/frontiercode; harness mini-swe-agent; Mean@5 |
| FrontierMath (Tiers 1-3) | 72.2% | Maxkimi-k3_max | Independent testIndependentEpoch AI ↗ | 17 Jul 2026 | — | FrontierMath Tiers 1-3 (v2), Epoch-run; stderr 2.7pt; data https://epoch.ai/data/benchmark_data.zip (frontiermath_tiers_1_3_v2.csv) |
| FrontierSWE | 81.2% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | Dominance score as of 2026-07-16 |
| FrontierSWE | 25.9% | Default | Independent testIndependentFrontierSWE ↗ | 1 Oct 2026 | proximus | FrontierSWE V2, mean@5 over 34 tasks (20h budget); ±11.8; $109.71/trial; 18.4h/trial; 'Per provider' view shows best entry per provider; Epoch notes runs use max reasoning effort |
| GDPval-AA v2 | 1686 | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | Cited from Artificial Analysis (2026-07-23) |
| GDPval-AA v2.1 | 1540 | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | 95% CI 1520.01-1559.96 |
| GPQA Diamond | 91.9% | Highkimi-k3_high | Independent testIndependentEpoch AI ↗ | 7 Aug 2026 | — | Epoch-run GPQA Diamond; stderr 1.9pt; data https://epoch.ai/data/benchmark_data.zip (gpqa_diamond.csv) |
| GPQA Diamond | 93.5% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run evaluation (AA slug kimi-k3) |
| GPQA Diamond | 93.1% | Maxkimi-k3_max | Independent testIndependentEpoch AI ↗ | 16 Jul 2026 | — | Epoch-run GPQA Diamond; stderr 1.5pt; data https://epoch.ai/data/benchmark_data.zip (gpqa_diamond.csv) |
| GPQA Diamond | 93.5% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | temperature 1.0, top_p 0.95 |
| GPQA Diamond | 92.9% | Default | Independent testIndependentVals.ai ↗ | 1 Sep 2026 | — | ±1.309 stderr; $0.088/test |
| Harvey LAB-AA | 94.6% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | criteria pass rate, AA-run |
| HLE Diamond | 22.2% | Defaultkimi-k3 | Independent testIndependentScale AI SEAL ↗ | 1 Oct 2026 | — | rank 10 (Scale rank accounts for CI); ±2.6; entry added 2026-04-22 |
| Humanity's Last Exam | 46.9% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run evaluation (AA slug kimi-k3) |
| Humanity's Last Exam | 43.5% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | HLE-Full without tools |
| Humanity's Last Exam (with tools) | 56.0% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | HLE-Full with general tools |
| IOI (Vals) | 48.9% | Maxreasoning_effort=max | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | — | ±9.82 stderr; $14.165/test |
| Kimi Code Bench v2 (internal) | 72.9% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | In-house 'Kimi Code Bench 2.0'; 73.7 with Claude Code harness |
| LegalBench (Vals) | 86.0% | Defaultvals id kimi/kimi-k3 | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | — | vals id kimi/kimi-k3; rank 16/149; ±0.388 stderr; $0.007066/test |
| LiveCodeBench | 87.2% | Maxreasoning_effort=max | Independent testIndependentVals.ai ↗ | 1 Sep 2026 | — | ±0.971 stderr; $0.104/test |
| LLM Creative Story-Writing Benchmark (Lech Mazur) | 2.5 | DefaultKimi K3 | Independent testIndependentLech Mazur (LLM Creative Story-Writing Benchmark) ↗ | 28 Sep 2026 | — | rank 9/56; Thurstone comparison score (centered at 0); est. win chance 82%; 95% bootstrap 2.383 to 2.527 |
| LMArena Code Arena (WebDev) | 1658 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | Code Arena | WebDev overall (agentic web-dev, raw); rank 12 (CI rank 9-13); 95% CI 1651-1664; 16246 votes |
| LMArena Text - Coding category | 1541 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | text arena coding category, style control on; rank 7 (CI rank 1-27); 95% CI 1533-1549; 7165 votes |
| LMArena Text - Creative Writing | 1460 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 26 (rank range 13-55); 95% CI 1450.9-1468.5; 5548 votes; style-controlled |
| LMArena Text - Expert | 1533 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 10 (rank range 1-31); 95% CI 1520.5-1544.6; 2568 votes; style-controlled |
| LMArena Text - Hard Prompts | 1518 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 9 (rank range 4-20); 95% CI 1512.0-1523.1; 17889 votes; style-controlled |
| LMArena Text - Instruction Following | 1487 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 13 (rank range 6-27); 95% CI 1480.5-1494.4; 9232 votes; style-controlled |
| LMArena Text - Longer Query | 1504 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 11 (rank range 6-25); 95% CI 1496.9-1510.0; 12240 votes; style-controlled |
| LMArena Text - Multi-Turn | 1500 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 9 (rank range 3-33); 95% CI 1490.8-1510.0; 4291 votes; style-controlled |
| LMArena Text - Non-English | 1476 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 19 (rank range 7-32); 95% CI 1470.8-1482.0; 17115 votes; style-controlled |
| LMArena Text - Occupational: Business, Management & Finance | 1488 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 14 (rank range 3-37); 95% CI 1479.3-1496.5; 5332 votes; style-controlled |
| LMArena Text - Occupational: Legal & Government | 1512 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 4 (rank range 1-31); 95% CI 1498.6-1524.9; 2206 votes; style-controlled |
| LMArena Text - Occupational: Writing, Literature & Language | 1476 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | rank 16 (rank range 8-36); 95% CI 1468.1-1483.7; 7298 votes; style-controlled |
| LMArena Text (overall) | 1488 | Maxkimi-k3-max | Independent testIndependentLMArena ↗ | 30 Sep 2026 | — | text overall, style control on; rank 17 (CI rank 8-27); 95% CI 1483-1493; 27719 votes |
| MCP Atlas | 82.3% | Maxkimi-k3 (max) | Independent testIndependentScale AI SEAL ↗ | 1 Oct 2026 | — | rank 2 (Scale rank accounts for CI); ±2.35; entry added 2026-07-20 |
| MCP Atlas | 84.2% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | 500-task public subset, 100-turn limit, Gemini 3.1 Pro judge |
| MCPMark Verified | 94.5% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | |
| MLS-Bench-Lite | 48.3% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | |
| MMMU-Pro (no tools) | 81.6% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | Without tools; 83.4 with Python |
| OSWorld-Verified | 84.8% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | |
| PostTrainBench v1.1 | 36.6% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Claude Code | Official Harbor implementation on H20 GPUs, avg of 3 runs |
| ProgramBench (fully resolved) | 77.8% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | Other models' scores from Vals AI |
| ProgramBench (fully resolved) | 2.0% | Default | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | mini-SWE-agent | ±0.992 stderr; $70.475/test; strict fully-resolved rate |
| SciCode | 59.5% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run evaluation (AA slug kimi-k3) |
| SciCode | 58.7% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | Cited by Moonshot from Artificial Analysis (2026-07-23) |
| SimpleBench | 60.7% | MaxKimi K3 (max) | Independent testIndependentSimpleBench ↗ | 17 Jul 2026 | — | AVG@5, temp 0.7; rank 33rd; human baseline 83.7%; data https://simple-bench.com/static/js/leaderboard-data.js |
| SimpleQA Verified | 50.6% | Maxkimi-k3_max | Independent testIndependentEpoch AI ↗ | 27 Aug 2026 | — | Epoch-run (no tools); ±1.58 stderr |
| SWE Atlas - Codebase QnA | 66.1% | DefaultKimi K3 | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | Kimi Code CLI | AA Coding Agent Index component; pass@1 avg of 3 attempts |
| SWE Marathon | 42.0% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Claude Code | H20-calibrated branch of official tasks (pre-v1.1, as of 2026-07-09) |
| SWE-Bench Pro V2 (full) | 97.7% | MaxKimi-K3 (mini-swe-agent) max | Independent testIndependentScale AI SEAL ↗ | 1 Oct 2026 | mini-SWE-agent | rank 3 (Scale rank accounts for CI); ±0.9; entry added 2026-09-22; SWE-Bench Pro V2 (642 tasks, locked protocol, released 2026-09-22) |
| SWE-Bench Pro V2 (hard) | 88.2% | MaxKimi-K3 (mini-swe-agent) max | Independent testIndependentScale AI SEAL ↗ | 1 Oct 2026 | mini-SWE-agent | rank 5 (Scale rank accounts for CI); ±0; entry added 2026-09-22; SWE-Bench Pro V2 (642 tasks, locked protocol, released 2026-09-22) |
| SWE-bench Verified | 93.4% | Default | Independent testIndependentVals.ai ↗ | 1 Sep 2026 | bash-only (single bash tool) agent | ±1.111 stderr; $0.760/test; Vals archived SWE-bench Verified on 2026-09-01 (saturated) |
| Terminal-Bench 2.1 | 85.0% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run evaluation (AA slug kimi-k3) |
| Terminal-Bench 2.1 | 88.3% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | Kimi Code | |
| Terminal-Bench 2.1 | 80.9% | Default | Independent testIndependentVals.ai ↗ | 27 Sep 2026 | Terminus 2 | ±0.649 stderr; $0.341/test |
| Terminal-Bench 4.0 | 19.7% | Maxreasoning_effort=max | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | mini-SWE-agent | ±3.499 stderr; $10.372/test |
| Terminal-Bench 4.0 | 12.6% | MaxKimi K3 (Max) | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | — | AA-run evaluation (AA slug kimi-k3) |
| Terminal-Bench 4.0 | 21.2% | DefaultKimi K3 | Independent testIndependentArtificial Analysis ↗ | 1 Oct 2026 | Kimi Code CLI | AA Coding Agent Index component run in the vendor agent harness; pass@1 avg of 3 attempts |
| Toolathlon-Verified | 76.5% | Maxreasoning_effort=max | Maker's own figureVendor-reportedMoonshot AI ↗ | 16 Jul 2026 | — | |
| Vals CorpFin v2 | 71.6% | Defaultvals id kimi/kimi-k3 | Independent testIndependentVals.ai ↗ | 12 Aug 2026 | — | vals id kimi/kimi-k3; rank 3/134; ±0.888 stderr; $0.063819/test |
| Vals Legal Research Bench | 44.2% | Defaultvals id kimi/kimi-k3 | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | — | vals id kimi/kimi-k3; rank 17/72; ±3.452 stderr; $3.222748/test |
| Vals Public Benefits Bench v1.1 | 68.3% | Maxreasoning_effort=max | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | — | vals id kimi/kimi-k3; rank 10/45; ±1.211 stderr; $1.03289/test |
| Vals TaxEval v2 | 75.7% | Defaultvals id kimi/kimi-k3 | Independent testIndependentVals.ai ↗ | 1 Sep 2026 | — | vals id kimi/kimi-k3; rank 12/145; ±0.841 stderr; $0.078156/test |
| Vals Vibe Code Bench | 85.0% | Default | Independent testIndependentVals.ai ↗ | 29 Sep 2026 | OpenHands | ±2.69 stderr; $17.590/test |