Kimi K3.1 Pro Weights Are Out: 1,561 GB and a License to Read

Moonshot AI published the full weights for Kimi K3.1 Pro on Hugging Face on November 4, 2026, twelve days after the model went live through its API. The promise made at launch was "by November 4" and Moonshot hit the date. That makes a 2.8-trillion-parameter, million-token-context model downloadable, modifiable, and self-hostable by anyone with the hardware. Almost nobody has the hardware.

Here is what actually shipped, what the license permits, and where independent evaluators say the model is strong and weak.

What is in the repository

I pulled the file listing from the Hugging Face API after release and added up the weights. The repository moonshotai/Kimi-K3.1-Pro contains 96 safetensors shards totaling 1,561,293,160,568 bytes, or about 1,561 GB. Kimi K3, released July 27, came in at roughly the same size, which is consistent with a post-trained model built on the same architecture.

The weights are quantized with compressed-tensors in mxfp4-pack-quantized format. That means 4-bit expert weights with MXFP8 activations, produced through quantization-aware training rather than post-hoc compression. There is no separate FP16 or BF16 release, and because the checkpoint is already near 4.25 bits per parameter, the usual trick of quantizing further to get a workstation-sized model mostly does not work here.

The model card lists these specifications:

• 2.8 trillion total parameters, 104 billion active per token
• 93 layers, one of them dense
• 69 KDA attention layers plus 24 gated DSA layers. Compared with Kimi K3, the full-attention gated MLA layers are replaced with gated DeepSeek Sparse Attention for cheaper long-context modeling
• 896 experts, 16 selected per token, plus 2 shared experts
• 160K vocabulary, 1,048,576-token context window
• A MoonViT-V2 vision encoder with 401 million parameters. Input is text and images

The "3T class" label refers to total parameters. What decides your bill and your GPU count is the 104 billion active figure, and conflating the two is how people misjudge both hosting cost and API economics.

The license is not OSI-approved, and the conditions matter at scale

The published LICENSE file reads like MIT through most of its length. It grants free use, copy, modify, merge, publish, distribute, sublicense, sell, and the right to run, deploy, fine-tune, and build derivative works. Two conditions sit on top of that grant.

Section 2 defines "Model as a Service" as giving a third party access to inference or fine-tuning, for example via an API, in a way that allows meaningful control over inputs, parameters, or training data. If you or an affiliate operate such a business and your aggregate revenue exceeds $20 million over any consecutive 12 months, you must sign a separate agreement with Moonshot before any commercial use. The definition explicitly excludes end-user products where the model is embedded inside a feature or harness, and mere relaying of requests to models someone else hosts.

Section 3 is an attribution clause: if a commercial product or service built on the model passes 100 million monthly active users or $20 million in monthly revenue, "Kimi K3.1-Pro" must be prominently displayed in the interface. Section 4 exempts internal use and access through Moonshot's official products or certified inference partners from both requirements.

In practice, a small team embedding the model in its own product operates close to an MIT grant. The clause bites inference resellers, hyperscalers, and neoclouds at scale. Hugging Face lists the license as other, not an SPDX identifier, so "open weights" is accurate and "open source" is not. Coverage that uses the two interchangeably is describing something other than what is in the file.

What it takes to run

You cannot fit 1,561 GB on one machine. You need aggregate accelerator memory above that number with room left for KV cache, activations, and communication buffers. Reasonable shapes, from published deployment guidance and the arithmetic the checkpoint imposes:

• Eight 288 GB accelerators (B300, GB300, MI350X class): 2,304 GB aggregate, about 740 GB of headroom over the weights
• Sixteen 180 GB accelerators (B200): 2,880 GB aggregate
• Sixteen 141 GB accelerators (H200): 2,256 GB aggregate
• Thirty-two 80 GB accelerators (H100): 2,560 GB aggregate

Eight H100s give you 640 GB and fall 921 GB short of the weights alone, before any runtime allocation. K3 had the same problem, and K3.1 Pro inherits it. The model card recommends multi-node serving through SGLang or vLLM. Storage is its own constraint: you need more than 1.5 TB free for the download and realistically 2 TB or more once extraction and cache are counted.

This is not a release for people running models on a desktop. It is a release for organizations with on-premises or dedicated clusters, typically for data-residency or sovereignty reasons.

How it actually performs

Moonshot's own comparison table puts Terminal-Bench 4.0 at 65.7, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. It reports SWE Marathon v1.1 at 51.7, HLE-Full with tools at 70.0, BrowseComp at 92.1, Automation Bench-AA at 73.3, and OSWorld V2.1 at 79.6.

Independent runs land in the same neighborhood. Vals AI measured 63.64% on Terminal-Bench 4.0, within two points of Opus 5.5 at 65.15%, and 62.52% on its Vals Index, which places the model in the top ten overall. Artificial Analysis recorded 55.1 on its Intelligence Index, fourth across everything it tracks, with 67.2% on Terminal-Bench 4.0, the best score it has measured.

The weaknesses are equally specific. Vals AI found K3.1 Pro is worse than Kimi K3 on legal and tax work: Legal Research Bench at 37.50% against K3's 46.15%, Tax Agent Bench at 24.87% against 33.37%, and HLAB-Vals-1.1 at 11.67% against 12.92%. Terminal work and code generation improved sharply generation over generation; specialist research regressed. A procurement decision made on one composite score will miss that.

Pricing is $2.00 per million input tokens and $10.00 per million output, with cached input at $0.20, down from $3.00 and $15.00 for K3. Moonshot says the rate is flat across the full million-token context. The Highspeed variant runs 2 to 3 times faster at double the price, $4.00 and $20.00.

How to decide

Do not treat weight availability as a mandate to self-host. If you do not already operate a GPU cluster and staff to run it, the API is still the correct answer after release. The practical value of open weights is narrower and more durable: independent audit, and the guarantee that if Moonshot changes pricing or terms, a competing provider can serve the same model.

Segment by workload rather than by headline. Terminal operations, code repair, and long-horizon agent tasks are measurably strong. Legal and tax research are weaker than the previous generation. Run your own evaluation on your own tasks before committing, and measure task completion rather than an aggregate index.

Read the license if you plan to resell inference. The $20 million revenue line decides whether you need a separate agreement. For internal use and embedded products, the terms are close to MIT. Commentators have described the terms as both "non-commercial" and "completely free"; neither matches the text.

Kimi K3 shipped weights 11 days after its API launch in July. K3.1 Pro shipped them 12 days after launch in October. That interval now looks like a policy rather than a delay. The next developments worth watching are community quantization progress and fresh benchmarks run on the released weights rather than the hosted endpoint, since those two results decide whether self-hosting becomes practical for teams below hyperscale.