Kimi K3.1 Pro Launches with Reasoning Mode Required; Full Weights Expected in November

Moonshot AI made Kimi K3.1 Pro available through services including Kimi Code on October 23 and also launched a Highspeed version. The company describes it as a three-trillion-class model that supports a one-million-token context window. As of the time of review, the full weights had not been released; a technical post said they would be published by November 4 at the latest. For now, the hosted service is what can be evaluated, while deployment costs and licensing terms await further documentation. Official launch notes, technical post

The technical approach builds on Kimi K3, with the main changes being more post-training compute and adjustments to the reinforcement learning framework, long-context inference, sandboxing, and rollout scheduling. The company reports a score of 65.7% on Terminal-Bench 4.0 using the Claude Code execution framework. Results of this kind are also affected by tools, the execution environment, and the reasoning budget, so they should not be taken as the model’s success rate across all coding tasks. Technical post and benchmark notes

External evaluations offer another perspective. On October 23, Artificial Analysis reported an Intelligence Index score of 55, up from 44 for K3. It measured 67% on Terminal-Bench 4.0, while AutomationBench-AA rose from 58% to 73%. These results support the view that agent workflows have improved, but they come from different tests than the company’s figures and should be considered separately. Artificial Analysis evaluation

The first change developers may encounter is API behavior. The Kimi Code documentation says K3.1 Pro supports thinking mode only, with low, high, and max settings and max as the default; requests that disable thinking are rejected. The company also recommends starting a new session when switching models to prevent existing reasoning content from affecting outputs and to improve cache hits. Converting protocols through agent tools may also change behavior. Model configuration documentation

Engineering teams should next measure task completion rates, total token usage, and end-to-end latency on their own codebases. The two-to-three-times speed increase for Highspeed is a company claim, and Kimi Code usage quotas are about twice those of the standard version. The one-million-token context window is also subject to membership-plan limits. When choosing a model, compare speed, quota usage, and the context window actually available, and evaluate self-hosting options once the weights are released. Model configuration documentation