What changes compared with K3
Moonshot AI has announced Kimi K3.1 Pro, the first model in the K3.1 series: according to the source, it is a post-trained version of Kimi K3 with much more compute dedicated to this phase. The company says it conducted joint reinforcement learning across hundreds of application domains, also working on RL frameworks, long-context inference, sandbox infrastructure, and rollout planning.
The base architecture remains that of K3, with Kimi Delta Attention and Attention Residuals, and activates 16 of 896 experts via a Stable LatentMoE framework. The model has 2.8 trillion total parameters and 104 billion active parameters, with a one-million-token context window; via API it accepts text, images, and video. It is not a model that can run on a single GPU: for K3, Moonshot specifies supernodes of at least 64 accelerators, with weights in MXFP4 and activations in MXFP8.
Availability, pricing, and the Highspeed version
K3.1 Pro is available on Kimi.ai, Kimi Work, Kimi Code, and via an OpenAI-compatible API endpoint, under the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The listed prices are $2 per million input tokens and $10 per million output tokens, with cache hits at $0.20; the Highspeed variant costs $4 and $20 and, according to the official blog cited by the source, runs two to three times faster. K3 was priced at $3 and $15.
The most delicate point concerns openness: the source describes K3.1 Pro as an open-weight model, but notes that the weights are not yet available and that release is expected in November 2026. Until then, access is exclusively through the paid API. Moonshot reportedly also removed cyber-related data from mid-training and RL ahead of the open release: on an unfiltered internal test, the success rate would have been 90.5% at pass@3, while the distributed model stands at 81.0%.
The claimed performance and the limits of verification
All results are self-reported, obtained according to Moonshot at maximum reasoning effort and predominantly with the Claude Code harness, and should be read as the company's claims. In coding, the model scores 65.7% on Terminal-Bench 4.0, just below the 66.4% attributed to Claude Opus 5.5 in Kimi's chart, while on SWE-Marathon it leads with 51.7 versus 46.3. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark, it rises to 61.2 from K3's 55.95.
The sharpest gap emerges in terminal-based scientific tasks: 38.1 on Terminal-Bench Science 0.1, versus 68.1 for GPT-6 Astra and 63.3 for Opus 5.5, again according to the company's chart. On BenchCAD with tools, the score is 92.7 (Opus 5.5 at 96.2, GPT-6 Astra at 95.9); without tools, 72.8 versus 73.0. In finance, Moonshot claims 64.4 on OfficeQA Pro v2 and 58.6 on Frontier Finance, with internal gains of 7.2 points in modeling and 11.6 in research compared with K3. On the internal Kimi Hallu Bench Hard test, statement-level hallucinations fall from 23.4% to 8.3%, versus 5.5% for GPT-6 Astra. The entire comparative picture remains to be independently verified, which the source itself describes as unconfirmed on competitors' pages.
