Kimi K3Open Weight
Released
Kimi K3 is an AI model released by Moonshot AI on Thursday, Jul 16 2026, 34 days after Kimi K2.7 Code. It is an open-weight model — the trained weights are available to download and run. It is a 2.8T parameter model with a 1M token context window. Benchmark results (shown below) cover DeepSWE 1.0, CheatBench, BullshitBench v2, DeepSWE 1.1, Next.js Evals, Supabase Evals, and 12 more.
Get Moonshot AI releases and AI news.
Every other Monday
API pricing
- InputWhat you pay for everything you send the model — your question, plus any documents or earlier conversation you include with it.
- $3.00
- Cached inputA reduced rate for text you send over and over. If every request starts with the same instructions or the same document, the provider keeps a copy ready and charges less to read it again.
- $0.30
- OutputWhat you pay for the text the model writes back. It is normally the dearer half: producing an answer costs more than reading one.
- $15.00
Available from
| Chutes | — | — | $1.85 | $9.50 | 1M | mxfp4 |
|---|---|---|---|---|---|---|
| Wafer | us | — | $0.70 | $10.00 | 1M | — |
| Decart | — | — | $2.34 | $11.70 | 1M | mxfp4 |
| Makora | — | — | $1.53 | $12.75 | 1M | fp4 |
| Phala | — | — | $2.55 | $12.75 | 1M | — |
| DigitalOcean | — | — | $2.55 | $12.95 | 1M | — |
| Parasail | — | — | $2.60 | $13.00 | 1M | fp4 |
| Sail Research | — | — | $0.84 | $13.50 | 1M | fp4 |
| Relace | — | — | $0.496 | $14.00 | 1M | fp4 |
| AkashML | — | — | $1.20 | $14.00 | 1M | fp4 |
| DeepInfra | — | — | $2.85 | $14.25 | 1M | mxfp4 |
| InferenceNet | — | — | $0.62 | $15.00 | 1M | fp4 |
| Parasail | — | fast | $3.00 | $15.00 | 1M | fp4 |
| BaseTen | — | — | $3.00 | $15.00 | 1M | fp8 |
| Fireworks | — | — | $3.00 | $15.00 | 1M | — |
| Modal | — | — | $3.00 | $15.00 | 1M | mxfp4 |
| Moonshot AI | — | — | $3.00 | $15.00 | 1M | mxfp4 |
| Together | — | — | $3.00 | $15.00 | 1M | — |
| Alibaba | — | — | $3.45 | $17.25 | 1M | — |
| InferenceNet | — | fast | $2.99 | $18.00 | 250K | fp4 |
| Morph | — | — | $0.67 | $20.00 | 1M | fp8 |
| Fireworks | — | fast | $4.50 | $22.50 | 1M | — |
| Fireworks | us | — | $4.50 | $22.50 | 1M | — |
| Wafer | — | — | $0.72 | $24.00 | 1M | — |
Benchmarks
Coding
DeepSWE 1.0Agentic coding — Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. Higher is better.
67.5%
#1Best published DeepSWE score of all tracked models
DeepSWE 1.1Agentic coding — Artificial Analysis' independent test of deep, agentic software-engineering work — the AI has to plan and carry out substantial coding tasks end to end. (Version 1.1 of the test.) Higher is better.
69%
#13 of 34Best: Muse Wonder Full · 79%
Next.js EvalsNext.js coding — Vercel's open eval of how well AI coding agents build and migrate real Next.js apps — measured as the share of tasks the agent completes successfully. Higher is better.
84%
#8 of 32Best: Claude Fable 5.1 · 97%
Supabase EvalsSupabase coding — Supabase's own open benchmark: a coding agent is dropped into a real Supabase project and asked to do real work — set up a schema, fix a broken security policy, debug an Edge Function — and every run is checked against a live Supabase stack. This is the headline number, where the agent has Supabase's own skills loaded, as most people building on Supabase would. The score is the share of scenarios it got right. Higher is better.
87%with skills
84.1%no skills
with skills
#4 of 11Best: Claude Opus 5 · 92.8%
no skills
#3 of 11Best: Claude Opus 5.5 · 91.3%
Terminal & CLI
Terminal-Bench 2.1Agentic terminal coding — Can the AI work in a command-line terminal — running commands and finishing technical setup tasks the way a developer would? Higher is better.
88.3%
#5 of 32Best: DeepSeek-V4.1-Flash · 90.6%
Agentic & tool use
MCP AtlasMulti-step tool use — Can the AI chain together many tools and steps to complete one bigger task, rather than doing just a single thing? Higher is better.
84.2%
#2 of 11Best: Muse Spark 1.1 · 88.1%
JobBenchProfessional tool use — Tests the AI on professional workplace tasks that require using real work tools — the kind of multi-step jobs an office worker handles. Higher is better.
52.9%
#8 of 10Best: Muse Wonder Full · 75.3%
Toolathlon-VerifiedPersonal tool use — Tests how well the AI uses everyday personal tools and apps to get things done — a human-checked version of Toolathlon. Higher is better.
73.2%
#6 of 9Best: GLM-5.3-Flash · 78.4%
BrowseCompWeb browsing — Can the AI browse the web and track down hard-to-find answers? Higher is better.
91.2%
#3 of 28Best: GPT-5.6 Sol · 92.2%
Reasoning & science
Humanity's Last ExamMultidisciplinary reasoning — Humanity's Last Exam — extremely hard expert questions across many subjects, written so you can't just look up the answer. “No tools” means the AI answers on its own. Higher is better.
43.5%no tools
56%with tools
no tools
#7 of 23Best: Claude Fable 5.1 · 60.9%
with tools
#16 of 48Best: Claude Opus 5.5 · 67.7%
ARC-AGI-2Abstract reasoning — Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better.via BenchLM
60.4%
#24 of 37Best: GPT-6 Astra · 95%
GPQA DiamondScience — Graduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better.
93.5%
#6 of 59Best: GPT-6 Astra · 96%
Knowledge work
GDPval-AA v2Knowledge work — economically valuable knowledge work (v2, re-based Elo)
1668
#9 of 20Best: Claude Opus 5 · 1861
Multimodal
CharXiv ReasoningChart reasoning — Can the AI read and reason about complex charts and figures, not just text? Higher is better.
84.8%
#5 of 16Best: Muse Spark · 88.9%
MMMU-ProMultimodal reasoning — A tougher version of MMMU — college-level questions that mix images, diagrams, and text together. Higher is better.
81.6%
#4 of 12Best: Gemini 3.5 Flash · 83.6%
Robustness
CheatBenchCheating rate — Measures how often AI agents try to cheat on difficult assignments, such as reading hidden answers, copying work or manipulating grading. The overall score gives equal weight to ten categories; the sycophancy category measures how far an agent shifts its beliefs toward a user's stated views. Scores describe each model in its tested agent setup, and task success is measured separately. Lower is better.via CheatBench
70%
#7 of 13Best: Claude Opus 5.5 · 11.2%
BullshitBench v2Nonsense detection — Given a confidently-worded but nonsensical prompt, does the AI spot that it makes no sense and push back — instead of playing along and inventing an answer? The score is how often it clearly called out the nonsense. Higher is better.
74%
#13 of 97Best: Claude Opus 4.8 · 95%
Community preference
threejsevalCommunity preference (Three.js) — Every model gets the same prompt — "the Eiffel Tower", "a glass fishbowl", "a robot arm picking toys into a box" — and builds a 3D scene in Three.js. Real people then see two scenes side by side, names hidden, and vote for the one they prefer. The votes become a chess-style Elo rating on threejseval.com, averaged across all the prompts. It measures whether the scene looks and moves right to a human eye, not whether the code passes a test. Higher is better.
1527
#12 of 27Best: Claude Opus 5.5 · 2074
Source: BenchLM, retrieved 28 September 2026 · Source: CheatBench, retrieved 25 September 2026. Other sources are identified on the linked benchmark pages.
Compare Kimi K3 with
About
Kimi K3, released July 16, 2026, was Moonshot AI's biggest swing yet: a 2.8-trillion-parameter mixture-of-experts model — billed as the world's first open 3T-class release — activating 16 of 896 experts per token, with a 1M-token context window and native vision built on the lab's Kimi Delta Attention architecture. At launch it scored 93.5% on GPQA Diamond, the strongest open-weight result on that benchmark published at the time, alongside 88.3% on Terminal-Bench 2.1.
Agentic work was the headline: 91.2% on BrowseComp — the best published score on this tracker at release — plus 56.0% on Humanity's Last Exam with tools and 84.2% on MCP Atlas. Moonshot's own framing placed K3 just behind Claude Fable 5 and GPT-5.6 Sol overall while undercutting both sharply on price, and the lab promised full open weights within days of the API launch — the same open-frontier playbook that began with K2 a year earlier, now at nearly triple the scale.
Frequently asked questions
Kimi K3 was released by Moonshot AI on Thursday, Jul 16 2026.