Kimi K3.1 Pro
Moonshot AI · Vision-language · MoE 2.8T (A104B) · 1049k context · Released 4 November 2026
Kimi K3.1 Pro is a sparse mixture-of-experts of around 2.8 trillion total parameters with about 104 billion active for any given token. It is natively multimodal, accepting text and images (via a 401-million-parameter MoonViT-V2 vision encoder) and producing text, with a context window of 1,048,576 tokens. Unusually, Moonshot trained it to run at low precision and distributes it that way: MXFP4 weights with MXFP8 activations, through quantisation-aware training. The published files are safetensors; there is no official GGUF, though community quantisations for llama.cpp, Ollama and LM Studio have appeared, and Moonshot points at vLLM and SGLang for serving. The API launched on 23 October 2026 and the open weights landed in early November. At roughly 1.4 to 1.5 terabytes of weights at its native 4-bit precision, it is a multi-node deployment: a 512GB unified-memory machine, which can hold a 2-bit build of the one-trillion-parameter Mistral Large 4, does not come close here.
Strengths
- Frontier-scale open-weight model that can be self-hosted with data kept in-house
- Natively multimodal, accepting images alongside text
- Context window just over one million tokens
- Sparse mixture-of-experts, so only about 104B of the ~2.8T parameters compute per token
- Permissive MIT-derived licence that behaves as open commercial for all but the largest resellers
Weaknesses
- Very large footprint, around 1.4 to 1.5TB at native 4-bit, so multi-node server hardware rather than a single machine
- Trained natively at 4-bit, so there is little headroom to compress it further without a steeper quality cost
- No official GGUF at release; running it locally relies on community quantisations
- Capability figures so far are Moonshot's own, and some comparisons against closed models have been questioned by outside testers
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| MXFP4 (as released) | ~1500GB | The native released format, MXFP4 weights produced by quantisation-aware training, roughly 1.4 to 1.5 terabytes of weights before context. This is multi-node server territory. Figure is approximate and rises with context and concurrency. |
| Community sub-4-bit (approx) | ~730GB | A rule-of-thumb for an aggressive 2-bit community quantisation of the 2.8T total, around 730GB. Even this remains a multi-GPU or very-large-memory deployment, and because the model was trained at 4-bit, going lower costs more quality than usual. Approximate; verify against a specific build. |
What you'd need to run this
Licence
Kimi K3.1 Pro License — read the licence
Kimi K3.1 Pro: common questions
- What hardware do I need to run Kimi K3.1 Pro?
- At its most compressed (Community sub-4-bit (approx)) it needs roughly 730GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is Kimi K3.1 Pro free for commercial use?
- Commercial use is permitted, but with conditions. Open weights under a bespoke "Kimi K3.1 Pro License", not an OSI licence, but derived from the MIT licence and permissive. Read the licence before relying on it at scale.
- Can I run Kimi K3.1 Pro on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is Kimi K3.1 Pro's context window?
- Kimi K3.1 Pro has a context window of 1,048,576 tokens, about 1049k.
Availability
Recommended for
- Organisations with multi-node infrastructure wanting a frontier-scale open multimodal model
- Long-context and long-document work up to about one million tokens
- Self-hosting where data must stay in-house and the model's size can be accommodated
Related models
Related guides
Catalogue entry last verified 9 November 2026. Specifications change; verify anything you are about to spend money on.