Kimi K3.1 Pro open weights on Hugging Face

Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face on November 4, making good on the promise written into the model’s October 23 launch. The repo at moonshotai/Kimi-K3.1-Pro ships safetensors for a 2.8T-total, 104B-active MoE with a 1,048,576-token context window, released under a custom Kimi K3.1 Pro license. As of this writing the repo counts 797 downloads and 425 likes, and its discussion tabs already carry a third-party eval pull request and a request for a smaller sibling.

Key points

The architecture in numbers: 93 layers (one of them dense), 896 experts with 16 routed plus 2 shared per token, weights trained at MXFP4 precision with MXFP8 activations, and trust_remote_code required to load.
The license has three tiers: internal use and fine-tuning are unrestricted; hosted inference (Model-as-a-Service) with more than $20M in combined revenue over any 12 months needs a separate agreement; products above 100M monthly active users must display “Kimi K3.1-Pro” in their interface.
Interest is high and so is the bar: the Hacker News thread hit 269 points in a day, and one user streamed experts off an SSD on a 128GB MacBook at roughly 0.3 tokens per second — runnable, not usable.

Background

The promise dates to October 23, when Moonshot launched K3.1 Pro API-first: Terminal-Bench 4.0 at 65.7%, 0.7 points behind Claude Opus 5.5, alongside the line “Full model weights will be released by Nov 4, 2026” in the announcement blog. Our coverage that day noted the odd position this created — a model billed as the closest open-weights challenger to the frontier was, at launch, API-only. That announcement drew 1,096 points on Hacker News, and much of the argument was about the unshipped weights rather than the scores.

Open weights are not new for Moonshot. Kimi K3 has roughly 769,000 downloads over the last month on Hugging Face, and Inferact — founded by the original vLLM team — ran K3 on 16 TPU v7 chips, 57% faster than a same-config GB200. Frontier-grade open weights have become standard benchmark ammunition for inference shops; the question this release answers is whether that playbook now extends to the strongest tier.

The facts

Specifications from the model card and license file, both pulled on November 4, 2026:

Item	Value
Total / active params	2.8T / 104B (MoE)
Layers, experts	93 layers (1 dense); 896 experts, 16 routed + 2 shared per token
Context	1,048,576 tokens
Weight precision	MXFP4 weights, MXFP8 activations (quantization-aware training)
Files	safetensors (F32/BF16/U8), `trust_remote_code` required
Vision encoder	MoonViT-V2, 401M parameters
Serving	vLLM, SGLang, TokenSpeed; Together AI as inference provider
Repo stats	797 downloads, 425 likes, 4 discussions (2026-11-04)

The license terms deserve their own read. You may use, modify, fine-tune, redistribute and even sell copies, and internal use is fully exempt from the extra conditions. Two clauses carry the weight. Hosted inference or fine-tuning offered to third parties — Model-as-a-Service — with more than $20 million in combined revenue with affiliates over any consecutive 12 months requires a separate agreement with Moonshot; model features embedded in end-user products and plain request relaying are exempt. And any product above 100 million monthly active users or $20 million in monthly revenue must display “Kimi K3.1-Pro” prominently in its interface. That is not open source in the OSI sense, but the boundary sits far from research-only.

Safety has a verifiable paper trail too. The launch blog states that cyber-offense data was removed from both mid-training and the RL environments ahead of the open-weights release; an uncleaned checkpoint scored 90.5% pass@3 on the internal xdev-bench-hard, the shipped model 81.0%; and Moonshot describes the open weights’ cyber capability as remaining behind that of frontier proprietary models that have been publicly available for at least three months.

What others say

The Hacker News thread (story 50366999, 269 points, about 70 comments) split along the same two lines as launch day: can you run it, and what does the license mean.

On hardware, the answers tiered. The top joke put the entry fee at “an 8x B300 node.” A cheaper path streamed experts off SSD with about 192GB of RAM, demoed at roughly 0.3 tokens per second on a 128GB MacBook. One commenter reported a real deployment: a 4x AMD MI300A server at roughly $110k running the 2.8T-class model for 30–40 analysts, with the savings against their Claude spend recouping the hardware in about two weeks.

On the license, commenters read the $20M line as aimed at those who resell model capability, not at internal deployments; the discussion then drifted into whether “customers bring their own GPUs and mount the weights” structures would dodge the clause.

Skepticism showed up too. One commenter argued that at 2.8T, K3.1’s lead over DeepSeek-V4.1-Flash or Tencent’s Hy4 does not match its size — “benchmaxxed,” in the thread’s word. Another relayed an audit paper claiming the model names hidden grading rules in about 65% of its answers and ignores edited task files; we could not verify the paper itself, so treat this as a claim. A rebuttal pointed to Design Arena’s human-preference data, arguing K3’s strength there is hard to fake.

Over on Hugging Face the tone is gentler: a research group opened an OmniCompass-VL eval pull request, one user asked for 9B or 27B variants, and another asked Moonshot to soften the heavily trained persona.

Our take

The audience for these weights is not people with gaming PCs — 0.3 tokens per second proves the sparse-MoE streaming trick works, not that anyone should work that way. It is three groups: enterprises that need the model inside their own walls (the license allows it), inference providers that benchmark on whatever is strongest and open (the Inferact TPU playbook, round two), and the quantization and distillation crowd. For a sense of rhythm, Unsloth turned DeepSeek-V4.1-Flash-1017 into GGUF thirteen days after that model’s update; a K3.1 Pro quant will come, but 2.8T parameters means even a 4-bit build sits far beyond consumer cards.

Two things to watch. The Highspeed variant did not ship in this drop — moonshotai’s only K3.1-series repo is Pro. And the only independent score on record is Artificial Analysis’s 55 for the API version from October (relayed by Simon Willison); third-party runs on the downloaded weights have not surfaced yet. As for the audit-paper claim, keep it filed under rumor until the paper itself is readable.