Kimi K3.1 Pro Brings an Open-Weight Model Within Reach of the Frontier

A 2.8T model near the top of the closed-model tables, with weights promised in 12 days.

Moonshot introduced Kimi K3.1 Pro on Friday as its most capable 3T-class open-weight model. It is a mixture-of-experts model with 2.8T total and 104B active parameters, a 1M-token context and multimodal input. Kimi K3.1 Pro and a Highspeed version, which Moonshot says is two to three times faster, are live on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Open weights are due by Nov 4.

Moonshot lists full-project coding of games in Unreal, Unity and Godot, parametric 3D modeling in Blender and CAD, and finance work with hallucination down by 65%. It says the model matches the strongest proprietary models across coding, agentic and knowledge benchmarks.

What the rankings say.

@kimmonismus counts wins over every listed US frontier model on 6 of 14 benchmarks, in Moonshot’s own evaluations. @MiaAI_lab reports No. 4 on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5, Claude Sonnet 5.5 and Muse Wonder Full, which makes it the No. 1 open-weight model. @simonw’s write-up gives the Artificial Analysis score as 55.

@simonw adds the pricing: $2.00 per million input tokens and $10 per million output, the same as GPT-6.1 Sol and half the price of Claude Opus 5.5. Highspeed costs twice as much, except for cached input. He ran his pelican test through Moonshot’s API with his llm tool at the low, medium and max reasoning levels. His verdict: the first time a soon-to-be open-weight model has competed this credibly with the current frontier.

The complaints.

@MiaAI_lab says her 5-hour limit on Kimi’s $39 plan was done exactly one minute into a prompt, and nothing was built. She recalls that the $39 plan included 5x agent credits at the K3 launch and now includes 2x, with 5x moved to the $99 plan. She calls DeepSeek V4.1 Flash 1017 the better model. @niklaslenz_ai says he is more excited about Qwen, GLM and DeepSeek, because Kimi is not cheaper here.

@kimmonismus calls it Fable 5.1 intelligence as a local model, once the weights ship. Running it at home is another matter. At 4 bits per weight, 2.8T parameters come to about 1.4 TB; eight 128 GB DGX Sparks hold 1 TB in total. Both figures are our own arithmetic, before any context cache.