FreshSynapse

AI trends & daily updates, in plain English

AI Models

Kimi K3.1 Pro's Weights Are Out: a 2.8-Trillion-Parameter Model You Can Now Download — If You Have a Data Centre

By FreshSynapse Editorial · · 4 min read

A blue-lit computer server unit in a data centre

Photo: Hippopx (public domain). Illustrative image.

What happened. On 4 November 2026, Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face, meeting the deadline it set when it launched the model on 23 October. The repository is moonshotai/Kimi-K3.1-Pro, it carries no access gate, and it holds 96 weight files totalling about 1.56 TB.

What the model is. The model card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model, post-trained from Kimi K3 and the first in the new K3.1 series. It lists 2.8 trillion total parameters with 104 billion active for each token, picked from 896 experts, a context window of 1,048,576 tokens, and text and image input through a 401-million-parameter MoonViT-V2 vision encoder. The weights are stored at 4-bit precision (MXFP4) after quantisation-aware training. Moonshot calls it its “most capable 3T-class open-weight model to date” — the company’s own claim.

How good it is — with the labels on. All benchmark figures below are Moonshot’s own unless stated otherwise, measured at maximum reasoning effort. On Terminal-Bench 4.0, which tests coding agents on real terminal tasks, Kimi K3.1 Pro scores 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for OpenAI’s GPT-6 Astra. On SWE Marathon it leads both, at 51.7% against 46.3% and 42.5%. On Terminal-Bench Science it trails badly, at 38.1% against 63.3% and 68.1% — the gains are not across the board. Two independent testers have run it: Artificial Analysis measured 67% on Terminal-Bench 4.0, and Vals AI gives it 63.64% on the same test and calls it the top open-weights model on its index, while noting it scored below Kimi K3 on legal research and tax tests.

The licence is not open source in the usual sense. The Kimi K3.1-Pro Licence is, apart from the model’s name, word for word the Kimi K3 licence. For most people it works like a permissive one: you may use, copy, change, fine-tune and sell the model, as long as you keep its copyright notice. Two conditions apply at scale: a business running a model-as-a-service operation whose group revenue passes $20 million over any 12 months must sign a separate agreement with Moonshot before commercial use, and a product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display “Kimi K3.1-Pro” in its interface.

What it takes to run. A data centre, or a rented piece of one. Moonshot’s card gives no memory or GPU figures but recommends vLLM, SGLang and TokenSpeed; vLLM’s simplest NVIDIA setup is eight GPUs — one B300 node or two GB300 trays — and the checkpoint is already quantised, with no smaller version offered. No GGUF build was found on Hugging Face on 4 November. One security note: vLLM’s command includes --trust-remote-code, which runs Python from the repository on your machine, so download from Moonshot’s own account rather than a copy — within an hour of the release, Hugging Face already listed re-uploads from other accounts. For most people the hosted route stays the practical one: the API price is unchanged at $2 per million input tokens and $10 per million output tokens, with cached input at $0.20, and a faster Highspeed version at $4 and $20.

What Moonshot took out. For the open-weight release, Moonshot says it removed cyber-related data from training. On its internal vulnerability-research test, K3.1 Pro solves 81.0% of problems, against 90.5% for the same training without the cuts, and on ExploitBench it scores 74.7, against 100 for GPT-6 Astra and 91.0 for Opus 5.5. Moonshot says it aims not to release open weights that would extend the absolute frontier of cyber capability.

The unconfirmed report. On 28 October, The Information reported that Moonshot and DeepSeek “relied heavily” on OpenAI’s GPT-6 Astra outputs to train their latest models, citing people with direct knowledge. That report remains unconfirmed: Moonshot credits K3.1 Pro’s gains to “substantially larger post-training compute” and does not mention distillation, and no public comment from any of the companies has been found. Separately, OpenAI said on 30 September that it had stopped a July campaign to extract its models’ hidden reasoning, which it attributed to individuals associated with Moonshot. If where a model’s training data came from matters to your company, ask the vendor before you build on it.

Sources

  • Moonshot AI
  • Kimi K3.1 Pro
  • open weights
  • Hugging Face
  • AI models
  • coding agents

Related articles