# Moonshot AI publishes the weights for Kimi K3.1 Pro, a 2.8-trillion-parameter mixture-of-experts model

The model activates 104 billion parameters per token and ships in a 4-bit format from quantization-aware training, and an independent index scores it 55, three points behind Claude Opus 5.5.

- Published: 2026-11-06T10:11:19.661Z
- Canonical: https://polylog.news/ai/2026-11-06/moonshot-ai-publishes-the-weights-for-kimi-k3-1-pro-a-2-8-tr
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [NVIDIA Blog](https://blogs.nvidia.com/blog/nvidia-life-chris-alexiuk/), [Startup Fortune](https://startupfortune.com/moonshot-ai-releases-full-weights-for-28-trillion-parameter-kimi-k31-pro/), [i10x](https://i10x.ai/news/moonshot-kimi-k3-1-pro-open-weights), [MindStudio](https://www.mindstudio.ai/blog/kimi-k3-1-pro-benchmarks-review)

Moonshot AI, the Beijing lab behind the Kimi models, [published the full weights for Kimi K3.1 Pro](https://startupfortune.com/moonshot-ai-releases-full-weights-for-28-trillion-parameter-kimi-k31-pro/) on Hugging Face. Most reports date the upload to November 4, which was the deadline Moonshot set when it launched the model through its API on October 23. The repository is public with no access form or waitlist, [according to secondary coverage](https://i10x.ai/news/moonshot-kimi-k3-1-pro-open-weights).

The model is a mixture-of-experts design with 2.8 trillion total parameters, 104 billion active per token, and a context window of about one million tokens. The weights use a 4-bit microscaling floating-point format (MXFP4) with 8-bit activations, produced by quantization-aware training. The download is about 1.56 TB across 96 safetensors shards. One report estimates that serving it takes about eight NVIDIA B300 GPUs, so in practice the weights are open to operators with data-center hardware.

Moonshot's own benchmark table, [as compiled by MindStudio](https://www.mindstudio.ai/blog/kimi-k3-1-pro-benchmarks-review), puts it at 70% on Humanity's Last Exam (HLE) with tools, against 67.7% for Claude Opus 5.5 and 57.2% for GPT-6 Astra. On Terminal-Bench 4.0 it scores 65.7 against Opus 5.5's 66.4. It trails substantially on ProgramBench, at 31.5 against 65.0. Moonshot ran its agentic coding rows inside the Claude Code harness. The more neutral number comes from Artificial Analysis, which scores the model 55, an 11-point gain over Kimi K3 and fourth overall behind Opus 5.5 at 58.

The license is Moonshot's own, not a standard open-source license. Reports say it is permissive for most users, but hosts earning more than $20 million a year need a separate agreement. API pricing is $2 per million input tokens and $10 per million output tokens.

NVIDIA has its own reasons to promote open weights. Its blog this week profiled [Chris Alexiuk](https://blogs.nvidia.com/blog/nvidia-life-chris-alexiuk/), a Toronto-based product research engineer who says open-source AI should be promoted "from the rooftops." Every large open checkpoint that enterprises host themselves increases demand for GPUs that enterprises buy directly, whoever trained the model.

## What this means

A downloadable model now sits within three index points of the best closed model, and that changes the price ceiling closed labs can charge for agentic coding. Its API price is roughly half Opus 5.5's reported $4 per million input tokens and $20 per million output tokens. The exposed parties are closed-model vendors with high per-token margins and Anthropic in particular, whose prospectus shows 80% consumption-based revenue just before its listing. Governments and firms outside the US gain an option that no US export decision can switch off once the weights are downloaded. That supports the desk's thesis that the US and China are the two AI superpowers. The factor that weighs against the thesis is that the weights run best on NVIDIA's top hardware, which keeps a point of US control at the serving layer. The consensus focuses on benchmark parity. The more durable effect is the $20 million revenue threshold in the license, which lets Moonshot charge large Western hosts while staying free for everyone else. The likelier outcome is that major US inference hosts list K3.1 Pro within weeks and cut their prices on closed-model-class workloads. That depends on the license terms being acceptable to US cloud legal teams. If large US hosts decline to serve it on compliance grounds, the read is wrong.

## What to watch

- Whether major US inference providers add Kimi K3.1 Pro to their catalogs, and at what price per million tokens. Wide listing would show that legal and compliance teams accept Chinese weights despite political risk.
- Independent agentic coding results run outside the Claude Code harness. These would show whether Moonshot's Terminal-Bench parity depends on the harness it chose.
- Any US policy move that treats open Chinese weights as a controlled import. That would extend export-control logic from chips to models running inside US infrastructure.