Eleven days of coverage quoted 5.6 TB for Kimi K3.1 Pro, and that number was wrong for the thing you can actually download. Moonshot AI published the weights on Hugging Face on November 4, 2026, and the repository is 1.56 TB across 96 safetensors shards. Same 2.8-trillion-parameter mixture-of-experts. Different precision, the one Moonshot trained it to serve in. That gap decides whether hardware you already own can hold it.
The second surprise is the licence. It isn’t Apache 2.0, it isn’t MIT, and it isn’t OSI-approved, despite a permissive opening paragraph. It has a revenue threshold that turns a free download into a negotiation for one specific kind of company.
What landed on Hugging Face
The repo is moonshotai/Kimi-K3.1-Pro. It is ungated and holds 111 files: 96 weight shards, plus the config, tokenizer, four Python files for the custom architecture, the model card and the licence. Total size is 1,561 GB.
- 2.8T total parameters, 104B active per token
- 93 layers, one dense, with 69 KDA layers and 24 Gated DeepSeek Sparse Attention layers
- 896 experts, 16 selected per token plus 2 shared
- 1,048,576-token context, unchanged from Kimi K3
- Text and image in, text out, using the 401M-parameter MoonViT-V2 vision encoder
- 160K vocabulary
Only the attention stack is new versus K3. Moonshot says this is a post-training release on the K3 base, with heavier joint reinforcement learning across hundreds of domains. The sparse attention swap targets long-context efficiency, not a capability jump.
The licence is the real fine print
The Kimi K3.1-Pro License grants the usual rights: use, modify, distribute, sublicense, sell, fine-tune. Then come conditions.
If you run a business selling third parties access to inference or fine-tuning, and group revenue passes $20 million over any 12 months, you need a separate agreement with Moonshot. Embedding the model in an end-user product doesn’t count as model-as-a-service. Neither does relaying requests to models someone else hosts. Separately, products with over 100 million monthly users or $20 million in monthly revenue must display “Kimi K3.1-Pro” in the interface. That’s branding, not payment. Internal use is exempt from both.
So internal deployments are free, and so is building your own product on top. The trigger is reselling raw model access at scale. If you planned a competing inference endpoint, take that clause to a lawyer. It’s structurally the same licence K3 shipped under, so existing legal reviews carry over. And it isn’t open source in the OSI sense: the metadata reads license: other. Outlets calling it modified MIT stopped reading after the first paragraph.
What it takes to run
The published vLLM recipe wants vLLM 0.34.0 or newer and a baseline of 8 GPUs in one tensor-parallel group, either two GB300 trays or one B300 node. The optional DCP8 long-context mode raises KV cache capacity to roughly 18.5 million tokens, against 2.6 million with FP8 KV on 8 B300s. That’s about seven times more.
Two consequences. An 8xH200 node is out, because 1.13 TB of HBM can’t hold 1.56 TB of weights before a single token of cache. And there is no consumer path. The quickstart also needs trust_remote_code=True, so the architecture isn’t in a released upstream library yet.
Does it matter?
For most teams, no, not directly. Realistically this is for organizations with B300-class hardware, data residency demands, or heavy enough volume to beat API pricing. Everyone else is better served by the hosted API, where Kimi K3.1 Pro scores 47 on Artificial Analysis intelligence, behind GPT-6.1 Sol and Claude Sonnet 5.5 at 55 each.
But the release matters as a signal. A near-frontier 2.8T model with a 1M context window is downloadable, with a licence that leaves almost everyone except inference resellers alone. That squeezes hosting margins and gives enterprises leverage in vendor talks. Expect third-party hosts to race to serve it, and expect the abliterated variants to follow quickly. Several already have.
