Kimi K3.1 Pro open weights are live: Moonshot's 2.8T model lands on Hugging Face on its promised date

Moonshot AI has published the open weights for Kimi K3.1 Pro on Hugging Face, meeting the November 4 deadline it set when the model launched on October 23. The repository, moonshotai/Kimi-K3.1-Pro, carries the full model card, a dedicated licence and safetensors weights. That turns a model that was hosted-only for its first 12 days into one anyone can download and run, given hardware on the scale of 2.8 trillion parameters.

The timing is the story as much as the files. Moonshot promised the full weights “by Nov 4” in its launch post, and this site covered that promise on launch day. One tracker that checked the repository at 06:15 UTC on November 4 still found it returning a 401. By the afternoon it was live, with the complete model card in place. The deadline, unusually for an open-weights promise, was met on the day.

What is in the release

The model card on Hugging Face is Moonshot’s own, and it leaves little to guesswork. Kimi K3.1 Pro is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, across 93 layers and 896 experts, 16 selected per token plus 2 shared experts. The context window is 1,048,576 tokens. A 401-million-parameter MoonViT-V2 vision encoder handles image input alongside text, and the weights are MXFP4 with MXFP8 activations, the product of quantization-aware training. For serving, Moonshot points to vLLM, SGLang and TokenSpeed, with recipes for each.

Kimi K3.1 Pro is post-trained from Kimi K3 with large-scale joint reinforcement learning, and it is the first model in Moonshot’s K3.1 series. A faster sibling, Kimi K3.1 Pro Highspeed, launched alongside it on October 23 and is described by Moonshot as 2 to 3 times faster; the weights release covers the Pro model. The weights ship under their own licence, named on the model card as the Kimi K3.1 Pro License. Read that licence before building anything commercial on the weights: the name is new, and its terms are the ones that count.

The numbers Moonshot is standing behind

All of the headline scores come from Moonshot’s own model card, run at maximum reasoning effort, so read them as the vendor’s case, not independent verification. On Terminal-Bench 4.0, Kimi K3.1 Pro scores 65.7, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra in the same table. On OSWorld V2.1, the computer-use test, it scores 79.6 against 81.8 for Opus 5.5. On GPQA-Diamond it reaches 94.2, and on HLE-Full, 51.1. The pattern is consistent: close to the best closed models on coding and agentic work, without quite passing them.

One score deserves a note of its own. On xdev-bench-hard, an internal cyber-vulnerability test, Kimi K3.1 Pro scores 81.0 at pass@3, below DeepSeek-V4.1-Flash-1017’s 85.7 in the same table. Moonshot has said it removed cybersecurity-related data during mid-training and reinforcement learning for the open-weight version, as this site reported in October. The weights now let anyone check what that filtering cost, and what it left intact.

What it will take to run it

Downloading the weights is the easy part. Running a 2.8-trillion-parameter model at full precision is not a laptop job, or even a single-workstation one: it is multi-GPU server hardware, the kind Moonshot’s recommended engines are built for. Most people who want to try the model will still do it the way they have since October 23, through the hosted API at $2 per million input tokens and $10 per million output tokens, with $0.20 per million cached input tokens. The weights matter for a different audience: labs and providers that want to serve the model themselves, fine-tune it, or evaluate it without going through Moonshot’s API. For a sense of what running models locally involves at a smaller scale, see this site’s hands-on with a tiny local AI harness — the principle is the same, the hardware is not.

What happens next is in other hands now: independent evaluations on the exact weights, quantised builds, and community ports to more engines. The October promise is kept, on the date given. The verdicts — on quality, on safety, and on whether a 2.8-trillion-parameter open model is practical outside a datacentre — are yet to come.