Kimi K3.1 Pro Open Weights Released by Moonshot AI

Key Points

• Moonshot AI released the full open weights of Kimi K3.1 Pro on Hugging Face on 4 November 2026, on the date it had promised at launch.
• The model has 2.8 trillion parameters in total, with 104 billion activated for each token, and a context window of about one million tokens.
• The download is about 1.56 terabytes across 96 files, so running it needs data-centre hardware — Moonshot’s own recipe uses eight high-end GPUs.
• The weights sit under Moonshot’s own Kimi K3.1 Pro License, not a standard open licence: large commercial API businesses face extra conditions.
• Moonshot’s own benchmark figures put the model close to Anthropic’s Claude Opus 5.5 on coding-agent tests, but those numbers have not yet been independently reproduced on the released weights.

Moonshot AI published the downloadable weights of its flagship Kimi K3.1 Pro model on 4 November, turning one of China’s most capable AI systems from a hosted service into something researchers and companies can inspect, adapt and run themselves.

What happened

Kimi K3.1 Pro first appeared on 23 October 2026, available through Moonshot’s Kimi apps, its Kimi Code coding tool and its API. From the start, the company promised something more: the full model weights would follow by 4 November. On that date, the files appeared on Hugging Face under the repository moonshotai/Kimi-K3.1-Pro, and Moonshot announced the release the same day.

The model card describes an open-weight, natively multimodal system built to act as an agent — software that can work through multi-step tasks rather than simply answer a single question. It is a mixture-of-experts design: 2.8 trillion parameters in total, spread across 896 experts, of which only 16 are switched on for any given token, plus two shared experts. That is why only 104 billion parameters are active at a time, keeping the running cost far below what the headline size suggests. It accepts text and images, holds a context of 1,048,576 tokens — roughly a whole library of documents in one conversation — and runs across 93 layers, including a newer sparse-attention design in 24 of them that Moonshot says handles long contexts more efficiently.

The download itself is a serious undertaking. The weights are stored in MXFP4, a compact 4-bit format produced by quantization-aware training, and still add up to about 1.56 terabytes in 96 files. Moonshot recommends the vLLM, SGLang or TokenSpeed engines to serve it, and its vLLM recipe runs the model across eight graphics processors — two Nvidia GB300 trays, one B300 node, or one AMD MI355X node. In plain terms, this is not a model for a laptop. It is for a lab, a cloud provider, or a company with a server room to spare.

The licence matters as much as the size. The Kimi K3.1 Pro License allows use, copying, modification and sale, but it is Moonshot’s own text, not MIT or Apache. A company selling model access that earns more than $20 million over twelve months must strike a separate deal with Moonshot before using the weights commercially, and a product with more than 100 million monthly users, or more than $20 million in monthly revenue, must display “Kimi K3.1-Pro” prominently. Internal use, and access through Moonshot’s own products and certified partners, are exempt from those conditions. Anyone planning a commercial product on these weights should read the licence first, not assume standard open-source terms.

There is also a deliberate safety cut. Moonshot says it found the model’s ability to find and exploit software flaws climbing as training scaled, so for the open-weight version it removed cyber-related data from mid-training and reinforcement learning. On its own internal exploit benchmark, a version trained without that cut solved 90.5% of tasks, against 81.0% for the released model. Moonshot’s stated aim, in its own words, is “not to release open weights that would extend the absolute frontier of cyber capability”. The benchmark figures here are Moonshot’s own, and the company does not claim the hosted API version and the downloadable weights are identical in this respect.

On price, the hosted version remains the cheap route for most users. Through Moonshot’s API, Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with a Highspeed variant at $4 and $20 that runs two to three times faster. That standard price is half the listed price of Claude Opus 5.5, the model Moonshot benchmarks itself against.

Why it matters

Open weights change who gets to check a model’s claims. A hosted model can only be tested through its API, on the provider’s terms. A downloadable model can be probed, fine-tuned, stripped down and rebuilt by anyone with the hardware — and its failures can be found by outsiders, not just announced by its maker. That is why this release, rather than the October launch, is the real test of Kimi K3.1 Pro: independent labs can now rerun Moonshot’s numbers on the actual files.

Those numbers, as reported by Moonshot, are striking. On Terminal-Bench 4.0, a test of coding agents on real terminal tasks, the company reports 65.7% for K3.1 Pro, against 66.4% for Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. Moonshot also reports 92.1 on BrowseComp and 94.2 on GPQA-Diamond. All of these are the maker’s own results, and none has yet been independently reproduced on the released weights — a caveat that should stay attached to every figure until outside tests land.

The release also sharpens a pattern running through this autumn’s AI launches. Chinese labs are now shipping frontier-class models as downloads, not just as services: Tencent’s Hy4 arrived the same week, and the gap between open and closed models keeps being measured in weeks rather than years. For smaller labs and startups, a model of this class that can be self-hosted is a building block; for the big US labs, it is another reason the lead they hold is a lead in deployment and scale, not a monopoly on capability.

The same week also saw AWS’s Semarak AI Malaysia programme put AI tools into the hands of government officers and small firms.

What happens next

Three things to watch. First, independent benchmarks: the first outside runs on the downloadable weights will show whether Moonshot’s coding-agent scores survive contact with other people’s test harnesses. Second, the ecosystem: smaller, compressed versions, community fine-tunes and cheaper serving recipes will decide whether this model stays a data-centre curiosity or becomes a practical tool. Third, the licence in practice: whether any large API provider takes up the weights under Moonshot’s extra conditions, or steers clear, will show how open this “open” release really is.

Sources: Hugging Face model card, moonshotai/Kimi-K3.1-Pro, 4 November 2026; Moonshot AI Kimi K3.1 Pro tech blog, 23 October 2026; MadRobot, updated 4 November 2026; AI Weekly, 4 November 2026; KimKJ China AI briefing, 5 November 2026. Benchmark figures are Moonshot’s own and await independent reproduction. Published: 5 November 2026, Kuala Lumpur time.