Kimi K3.1 Pro: open weights by November 4

On October 23 Moonshot AI introduced Kimi K3.1 Pro, the first model in the K3.1 series. Today it is only available on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Moonshot has promised the full weights by November 4, 2026. The SophyAI catalog already has its card, and on that day we will try to run it.

What model it is

K3.1 Pro is not trained from scratch. It is a post-trained Kimi K3, with much more reinforcement learning compute spread over hundreds of domains. The architecture is still K3's:

• 2.8 trillion total parameters, 104 billion active per token (MoE, 16 of 896 experts);
• a 1M-token context window (1,048,576);
• text, image and video input, text output;
• reasoning always on, with low, high or max effort.

It ships alongside K3.1 Pro Highspeed, which Moonshot says is 2–3× faster. For now Highspeed is API-only.

The claimed numbers

These figures come from Moonshot unless stated otherwise:

• Terminal-Bench 4.0: 65.7%, close to Claude Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%).
• BenchCAD with tools: 92.7%. It measures 3D reconstruction through CAD code.
• OfficeQA Pro v2: 64.4, which Moonshot calls state of the art.
• Claim-level hallucination rate on its internal Hard bench: from 23.4% for K3 down to 8.3%.

Artificial Analysis, an independent source, scores it 55 on its Intelligence Index. Only three proprietary models are ahead of it. Once the weights are out, it would be the highest-scoring open-weights model.

One caveat: Moonshot says it removed offensive-cyber data from the training of the open weights. The released weights may not fully match the model served through the API.

What it takes to run

If the architecture really is K3's, the requirements are the same. You need 8 B300 or B200 GPUs, at least 512 GB of RAM and 3 TB of NVMe disk. In MXFP4 the checkpoint should be around 1.7 TB. The engine is vLLM, using the Kimi K3 image. This is a cloud-session model, not a Mac one.

November 4

The plan is to reuse the kit we already run for Kimi K3 on 8× B300. We made the weights repository and the served model name configurable. As soon as the Hugging Face repository has the files, we will do four things:

1. read the license, which has not been announced yet;
2. download the weights and start vLLM;
3. measure tokens per second and hourly cost, with and without DSpark speculative decoding;
4. update the recipe and make it launchable from the portal.

Until then the k31pro-vllm-b300-upcoming card is visible but cannot be launched. If you want to try the model today, use the Kimi API. On SophyAI, Kimi K3 on 8× B300 is still available.

← All posts