Moonshot AI · 2026-11-02 · major
Kimi K3.1 Pro open weights — Moonshot's 2.8T model lands on Hugging Face
Kimi K3.1 Pro weights are now on Hugging Face: a 2.8T-parameter MoE with 104B active, 1M-token context and text plus image input, shipped in MXFP4. The custom license is MIT-style with limits for large MaaS businesses.
Moonshot keeps its promise: the full Kimi K3.1 Pro weights are now public on Hugging Face.
Key specs
| Parameters | 2.8T |
|---|---|
| Active params | 104B |
Quick facts
| Maker | Moonshot AI |
|---|---|
| Parameters | 2.8T total, 104B active |
| Context window | 1M tokens |
| Input | Text and image |
| Weights format | MXFP4 weights, MXFP8 activations |
| License | Kimi K3.1-Pro License (MIT-style, custom) |
| Serving | vLLM, SGLang, TokenSpeed |
Benchmarks
| Kimi K3.1 Pro | 65.7% | |
|---|---|---|
| Claude Opus 5.5 | 66.4% | |
| GPT-6 Astra | 58.2% | |
| Claude Fable 5.1 | 57.9% | |
| GLM-5.3 | 41.8% |
| Kimi K3.1 Pro | 70% | |
|---|---|---|
| Claude Opus 5.5 | 67.7% | |
| Claude Fable 5.1 | 65.6% | |
| GLM-5.3 | 62.5% | |
| GPT-6 Astra | 57.2% |
| Kimi K3.1 Pro | 51.7% | |
|---|---|---|
| Claude Opus 5.5 | 46.3% | |
| Claude Fable 5.1 | 45.6% | |
| GPT-6 Astra | 42.5% | |
| GLM-5.3 | 42.5% |
What is it?
Kimi K3.1 Pro's full weights went up on Hugging Face as moonshotai/Kimi-K3.1-Pro, meeting the November 4 date Moonshot AI set at the October 23 launch. The release is a 2.8T-parameter Mixture-of-Experts model with 104B parameters active per token, a 1M-token context and text plus image input.
How does it work?
Each token picks 16 of 896 experts, plus 2 shared experts, across 93 layers. Attention mixes 69 Kimi Delta Attention layers with 24 Gated DSA layers, which replace the full-attention Gated MLA layers in Kimi K3. A 401M MoonViT-V2 encoder handles images, and quantization-aware training ships the weights in MXFP4.
Why does it matter?
Teams that self-host can now run a model that, by Moonshot's own numbers, sits close to Claude Opus 5.5 on agentic coding and tool use. The license allows internal use and most products, with extra terms only for large model-as-a-service providers and very large apps.
Who is it for?
teams self-hosting frontier open models
Frequently asked questions
- Can I use Kimi K3.1 Pro weights commercially?
- Kimi K3.1 Pro ships under the Kimi K3.1-Pro License, which grants MIT-style rights to use, modify and fine-tune the weights. A model-as-a-service business with more than $20M revenue over any 12 months must sign a separate agreement with Moonshot AI, and products with over 100M monthly users or $20M monthly revenue must show the Kimi K3.1-Pro name.
- How do I serve Kimi K3.1 Pro myself?
- Moonshot's Kimi K3.1 Pro model card recommends vLLM, SGLang or TokenSpeed. The simplest start is pip install vllm and then vllm serve moonshotai/Kimi-K3.1-Pro. SGLang and Docker Model Runner commands are also listed. With 2.8T parameters in MXFP4, the model needs a multi-GPU server, not a laptop.
- How does Kimi K3.1 Pro compare to Claude Opus 5.5?
- In Moonshot's own model-card table, Kimi K3.1 Pro leads Claude Opus 5.5 on SWE Marathon v1.1 (51.7 vs 46.3), HLE-Full with tools (70.0 vs 67.7) and BrowseComp (92.1 vs 90.8). Opus 5.5 is ahead on Terminal-Bench 4.0 (66.4 vs 65.7), OSWorld V2.1, MMMU-Pro and Program Bench (65.0 vs 31.5). All scores are self-reported.
- Can I use Kimi K3.1 Pro without my own GPUs?
- Yes. Besides the Hugging Face weights, Moonshot's Kimi K3.1 Pro model card points to an OpenAI- and Anthropic-compatible API at platform.kimi.ai, where the model name is kimi-k3.1-pro. The model also runs in the Kimi chat app at kimi.ai, so a team can test it before deciding to self-host the 2.8T-parameter weights.
Try it
vllm serve moonshotai/Kimi-K3.1-Pro