OnPrem World地端世界

Models Published:

Moonshot releases Kimi K3.1 Pro weights: a 2.8-trillion-parameter model with a 1M-token context

Illustration: Moonshot releases Kimi K3.1 Pro weights: a 2.8-trillion-parameter model with a 1M-token context
Illustration

What happened

Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face on November 2. It is a natively multimodal (text and image) Mixture-of-Experts model with 2.8 trillion total and 104 billion active parameters, 896 experts with 16 selected per token, and a context window of 1,048,576 tokens. The weights use MXFP4 with MXFP8 activations, trained quantization-aware, and Moonshot recommends vLLM, SGLang or TokenSpeed to run it. It ships under the Kimi K3.1 Pro License, which is permissive but adds conditions for large Model-as-a-Service operators and very large consumer products, and exempts purely internal use.

Why it matters locally: It is one of the largest models anyone can download and host, and internal company use is exempt from its extra license conditions, but the files alone are about 1.56 TB, so it is a multi-GPU server model, not a desktop one.

Three ways to read it

For the market

In Moonshot's own tables, K3.1 Pro scores 65.7 on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2 and just behind Claude Opus 5.5 at 66.4, so open weights keep closing in on the top closed models, at least by the vendor's numbers. The license clause aimed at Model-as-a-Service businesses with more than US$20 million in revenue over 12 months suggests Moonshot, like Mistral, wants free company adoption while keeping big hosting rivals in a separate deal.

In plain words

A Chinese AI company has made its most capable model free to download. It is so large that it needs a room-sized server rather than an office PC, so most small companies will try it through Kimi's own website or API first.

For practitioners

Plan for about 1.56 TB of weight files; the vLLM recipe lists vLLM 0.34.0 or newer and an 8-GPU tensor-parallel group, such as one B300 node, two GB300 trays or one MI355X node, with the kimi_k3 tool-call and reasoning parsers. For multi-turn chats and tool calls, pass the full assistant message back, including reasoning_content and tool_calls, as the model card requires. Reasoning effort supports low, high and max, with max as the default, so test low for latency-sensitive internal tools.

Sprout's insight

Sprout compares this with other brands and with where real applications are heading. Free for members.

Sign in free to read it →

Subscribe to Sprout's weekly

One email a week: the top stories and a taste of Sprout’s insights. Unsubscribe any time.

Source

Read the official source: moonshotai/Kimi-K3.1-Pro model card (Hugging Face) ↗

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.

More news

All news →