Kimi
by Moonshot AI · Model · Language & code models
strong
The leading open-weights model for agents and web development, but enormous to run.
#1 agents (open weights)#1 documents (open weights)#1 web development (open weights)
#1 on Arena agent among open weights
AccessOpen weightsNeeds~1.56 TB; smallest build ~610 GBDownloadabout 1.56 TB full weights; one GGUF needs about 610 GBSize2.78T parametersLicenseKimi K3.1 Pro License
Why we say so
Arena ranks K3.1 first among open weights on agents and web development, and Artificial Analysis places it above Fable 5.1 and GPT-6 at about $2 per task. It also beat Qwen 4 Max on Terminal-Bench (67% versus 48%). The claim of near-frontier performance is Moonshot's own, and the weights are 1.56 TB, so most people will use an API or a heavily compressed quantization.
Versions
| Version | Released | Rank |
|---|---|---|
| Kimi K3.1LATEST strong Top open-weights model on Arena agents and web development | Oct 25 | #1 agents (open weights)#1 web development (open weights) |
| Kimi K3 | Jul 16 | #3 chat (open weights) |
| Kimi K2.6 promising First among open weights on Arena documents, second on vision | #1 documents (open weights)#2 image understanding (open weights) | |
| Kimi K2.5 |
Evidence
- Nov 18vLLM reports up to 30% higher acceptance length from a sliding-window draft for Kimi K3.1 Pro
- Nov 15#1 on Arena webdev among open weights (1684)
- Nov 11#3 on Arena text among open weights (1504)
- Nov 11#1 on Arena webdev among open weights (1687)
- Nov 9#1 on Arena document among open weights (1450)
Show all 39
- Nov 9#2 on Arena vision among open weights (1282)
- Nov 8
- Nov 8Full open weights released: 2.8T total, 104B active MoE, 1.56 TB; described as best open model, near top closed models
- Nov 6Took #1 on Arena webdev among open weights
- Oct 25Kimi K3.1 Pro released: 3T parameters, 1M context, on par with frontier models; open weights promised November 4
- Oct 25Ranks above Claude Fable 5.1, GPT-6 and Gemini 4 on Artificial Analysis index at about $2 per task
- Oct 25Highest Terminal-Bench score at lower cost per task than GPT and Claude, per vendor chart
- Oct 19#2 on Arena vision among open weights (1282)
- Oct 16Took #1 on Arena webdev among open weights
- Sep 272.78T-parameter MoE runs at about 1 tok/s on an M5 Max MacBook Pro (128 GB) with experts streamed from four SSDs
- Sep 13#1 on Arena document among open weights (1451)
- Sep 9Reported in the stolen-thoughts paper to get much closer to Claude 4.8 answers when prefilled with its reasoning, hinting at distillation
- Aug 28Found unpatched Mali GPU CVE-2022-38181 on the tablet's firmware, the starting point for the root exploit
- Aug 22Kimi K3 open weights reached 2,930 (52.2% gap closed) with Prime Agent harness; 2,974 with kimi-code, near Opus 5
- Aug 9Trained with full thinking history preserved; harnesses like Hermes that resend reasoning can destabilise quality, so Moonshot recommends its own harness
- Aug 4Fixed all bugs in about 8-9 minutes with ~17 live assertions, the most thorough checks, but left a stray booking uncleaned
- Aug 2Full weights released on Hugging Face on July 27 as promised: 2.8T MoE, 104B active, native vision (MoonViT)
- Jul 25Kebab prompt took 1h16m, about three times Opus 5's time; cheaper per token but similar total cost expected; thinner execution
- Jul 20Map-image geography and multilingual task: correct Netherlands ID, Dutch/Arabic/Portuguese lines, about 40 seconds with short reasoning
- Jul 19Output price a bit over three times cheaper than Fable 5; per launch posts K3 leads on Terminal-Bench, AutomationBench, GPQA, while Fable leads others
- Jul 19Beats Opus 4.8 on KernelBench CUDA; on KernelBench hard it is below Fable 5 but above Opus 4.8
- Jul 19Autonomously designed and optimized a small inference chip for a smaller model over 48 hours, in simulation
- Jul 17Moonshot launched Kimi K3, a 2.8T-parameter model with 1M-token context and native vision; API live, weights not yet public
- Jul 17Moonshot's own blog admits a noticeable gap to Claude and GPT frontier models; claims it beats Fable and Sol were refuted
- Jul 17Leads open models on Artificial Analysis index, about two points behind the best GPT, ahead of GLM 5.2
- Jul 17DeepSWE close to Fable 5 and GPT 5.6, over 20 points above Opus 4.8 and the next best open model
- Jul 17Cost-versus-performance charts place K3 as more cost-efficient per task than Claude Mythos and Opus
- Jul 17Ranked #1 on front-end code arena with 1679 points, ahead of Fable 5; #1 in six of seven front-end domains, #2 in gaming
- Jul 17Liquid splash simulator with webcam hand tracking built from scratch in one prompt; ran 30 minutes, 2M input tokens, 46K output
- Jul 17On the same prompt, Kimi K3 gave the richest simulation, live network animation, grounded assertions; reviewer rated it best, above Fable 5
- Jul 16Reported top on ProgramBench and SWE Marathon, beating Opus, both GPT models and GLM
- Jul 16Roughly tied with GPT Sol on Terminal-Bench, ahead of Opus and Fable
- Apr 22Took #1 on Arena document among open weights
- Jan 29Took #1 on Arena vision among open weights
Cost to run
Alternatives
Built on Kimi