Kimi introduces K3.1 Pro, a 3T-class model with open weights planned

What happened

Kimi introduced Kimi K3.1 Pro with 1M-token context, 2.8T total / 104B active params (MoE), and multimodal capabilities, claiming hallucinations are down by 65% in finance work. K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on http://Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with open weights planned by Nov 4, 2026. API: http://platform.kimi.ai Tech blog: http://kimi.ai/blog/kimi-k3-1-pro

Read next

• More changes in today’s briefingOlmoEarth v1.3 releases global satellite embeddings and a browser-based mapping explorer

Reported by 2 sources

The story so far

Kimi launched K3.1 Pro and K3.1 Pro Highspeed on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with model weights promised by November 4, 2026. Kimi describes K3.1 Pro as a multimodal mixture-of-experts model with a 1 million-token context window, 2.8 trillion total parameters and 104 billion active parameters. The company says the Highspeed variant is 2–3 times faster and claims K3.1 Pro reduces hallucinations in finance work by 65%. Nathan Lambert identified K3.1 Pro as the previously tested K3.1 and said its benchmark scores appeared strongest among current open-weight models, conditional on the promised weights release. Earlier, Leifeng.com reported that testers found strong Three.js graphics and interactive 3D results from the unreleased K3.1, alongside an awkward-looking giraffe animation. One comparison put its cost at about one-fourteenth that of GPT-6 Astra. Separately, Lambert said K3.1 costs less per token than K3, but higher token use on some benchmarks could leave the cost per answer similar.