Kimi K3Open Weight
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, Next.js Evals, Supabase Evals, BrowseComp, Arena Elo (Code), BullshitBench v2, and 11 more.
Benchmarks
Agentic coding
DeepSWE 1.0Artificial 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
Next.js coding
Next.js EvalsVercel'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.
92%
#1Best published Next.js score of all tracked models
Supabase coding
Supabase EvalsSupabase'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.
94.7%
with skills
100%
no skills
with skills
#2 of 5Best: GPT-5.6 Sol · 100%
no skills
#1Best published Supabase score of all tracked models
Web browsing
BrowseCompCan the AI browse the web and track down hard-to-find answers? Higher is better.
91.2%
#1Best published BrowseComp score of all tracked models
Community preference (code)
Arena Elo (Code)Like the text arena, but people vote on which AI writes better code. The votes become a chess-style Elo rating on arena.ai. Higher is better.
1679
#1Best published Arena Code score of all tracked models
Nonsense detection
BullshitBench v2Given 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.
73%
#9 of 67Best: Claude Opus 4.8 · 95%
Agentic coding
DeepSWE 1.1Artificial 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%
#3 of 14Best: GPT-5.6 Sol · 73%
Agentic terminal coding
Terminal-Bench 2.1Can 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%
#2 of 21Best: GPT-5.6 Sol · 88.8%
Multi-step tool use
MCP AtlasCan 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 10Best: Muse Spark 1.1 · 88.1%
Professional tool use
JobBenchTests 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%
#3 of 4Best: Muse Spark 1.1 · 54.7%
Personal tool use
Toolathlon-VerifiedTests 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%
#2 of 4Best: Muse Spark 1.1 · 75.6%
Multidisciplinary reasoning
Humanity's Last ExamHumanity'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
#5 of 14Best: Claude Opus 5 · 56.3%
with tools
#8 of 17Best: Claude Opus 5 · 64.7%
Science
GPQA DiamondGraduate-level science questions in biology, physics, and chemistry — hard enough that subject-matter PhDs score around 65%. Higher is better.
93.5%
#5 of 46Best: GPT-5.4-Pro · 94.4%
Knowledge work
GDPval-AA v2economically valuable knowledge work (v2, re-based Elo)
1668
#4 of 11Best: Claude Opus 5 · 1861
Chart reasoning
CharXiv ReasoningCan the AI read and reason about complex charts and figures, not just text? Higher is better.
84.8%
#4 of 12Best: Muse Spark · 88.9%
Multimodal reasoning
MMMU-ProA tougher version of MMMU — college-level questions that mix images, diagrams, and text together. Higher is better.
81.6%
#2 of 9Best: Gemini 3.5 Flash · 83.6%
Community preference
Arena Elo (Text)Real people chat with two anonymous AIs side by side and vote for the answer they prefer. Votes become a chess-style Elo rating on arena.ai — it measures which AI people actually like, not test scores. Higher is better.
1486
#7 of 17Best: Claude Fable 5 · 1509
About Kimi K3
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.
Kimi K3 — frequently asked questions
- When was Kimi K3 released?
- Kimi K3 was released by Moonshot AI on Thursday, Jul 16 2026.
- Who made Kimi K3?
- Kimi K3 was built by Moonshot AI. Chinese AI company building the Kimi family of models. Founded in 2023, known for long-context and efficient MoE architectures.
- What benchmark scores did Kimi K3 get?
- Kimi K3 reports 19 tracked benchmark scores — BullshitBench v2: 73%; DeepSWE 1.1: 69%; DeepSWE 1.0: 67.5%; Next.js Evals: 92%; Supabase Evals (with skills): 94.7%; Supabase Evals (no skills): 100%; Terminal-Bench 2.1: 88.3%; MCP Atlas: 84.2%; JobBench: 52.9%; Toolathlon-Verified: 73.2%; BrowseComp: 91.2%; Humanity's Last Exam (no tools): 43.5%; Humanity's Last Exam (with tools): 56%; GPQA Diamond: 93.5%; GDPval-AA v2: 1668; CharXiv Reasoning: 84.8%; MMMU-Pro: 81.6%; Arena Elo (Text): 1486; Arena Elo (Code): 1679. Scores are the figures published at release by Moonshot AI. It holds the best score among all models tracked here on DeepSWE 1.0, Supabase Evals (no skills), BrowseComp and Arena Elo (Code).
- What is the context window of Kimi K3?
- Kimi K3 has a context window of 1M. That is the maximum amount of input plus output the model can hold in a single request.
- How many parameters does Kimi K3 have?
- Kimi K3 is reported at 2.8T parameters.
- Is Kimi K3 open source?
- Partly. Kimi K3 is an open-weight model: the trained weights are free to download and run locally or on your own infrastructure, but the training data and code are not fully released and the license may restrict some commercial uses. It is not open source in the strict sense.
- What came before and after Kimi K3?
- Moonshot AI's previous tracked release was Kimi K2.7 Code on Jun 12 2026, 34 days earlier. It is the most recent Moonshot AI model tracked on AI Release Tracker.