Kimi K3.1 Pro and K3.1 Pro Highspeed
kimi.ai
October 23, 2026
15 min read
66/100
Summary
Kimi has introduced Kimi K3.1 Pro, a post-trained successor to Kimi K3 and its most capable 3-trillion-parameter-class open-weight model, with a one-million-token context window. Kimi K3.1 Pro Highspeed is also available, offering two to three times the standard model’s speed. Both models are available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API; Kimi says it will release the full weights by November 4, 2026. Kimi reports that K3.1 Pro scored 65.7% on Terminal-Bench 4.0, compared with 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5, and 61.2% on its internal Kimi Code Bench 3.0 Long-horizon, up from K3’s 55.95%. It reported a 92.7% tool-enabled score on BenchCAD and an OfficeQA Pro v2 score of 64.4. Kimi also said its internal stress test measured a claim-level hallucination rate of 8.3%, down from 23.4% for K3; those figures are not daily-use error rates. For the open-weight release, Kimi removed cyber-related data from mid-training and filtered cyber tasks from reinforcement-learning environments. On its xdev-bench-hard vulnerability benchmark, Kimi said a checkpoint without those measures achieved 90.5% pass@3, while the final model achieved 81.0%. Kimi said the controls reduced cyber capability while retaining general coding usefulness.
Key Takeaways
- Kimi K3.1 Pro is a 3T-class open-weight model with a one-million-token context window, and Kimi plans to release its full weights by November 4, 2026.
- Kimi reported a 65.7% Terminal-Bench 4.0 score for K3.1 Pro, above GPT-6 Astra’s reported 58.2% and below Claude Opus 5.5’s reported 66.4%.
- K3.1 Pro Highspeed is available alongside the standard model and is designed to run two to three times faster, according to Kimi.
- Kimi reported that K3.1 Pro’s internal factual-accuracy stress test produced an 8.3% claim-level hallucination rate, versus 23.4% for K3.
- Kimi removed cyber-related training data and reinforcement-learning tasks for the open-weight model; its xdev-bench-hard result fell from 90.5% pass@3 without those controls to 81.0% in the final model.
What the discussion said
Commenters treated K3.1 less as a simple release announcement than as a test of whether Moonshot can deliver frontier-class capability cheaply without shipping an open cyber weapon. The published scores and reported generation examples sparked real excitement: several readers saw a major leap, while others wanted the complete benchmark table to load before treating the claims as settled. Users also noted that Kimi has useful, uneven strengths alongside competing models, including persistence when navigating hostile websites, but argued that aggregate leaderboards do not answer the practical question of which model wins at UI work, coding, research, or agentic browsing. The sharpest conversation centered on its cyber policy. Some readers saw deliberately limiting offensive cyber knowledge during training as more robust than relying on post-hoc refusals: a released model cannot be jailbroken into expertise it never learned. Others read the claim that stronger internal systems exist as unverifiable marketing or politically motivated withholding. Cost was another clear bright spot: the new API is cheaper than K3 despite a slightly larger parameter count, strengthening the impression that capable models are becoming less expensive. That accessibility was tempered by reports that paid chat subscriptions exhaust their multi-hour quotas almost immediately, making the consumer product hard to use for sustained AI work.
Where opinion split
The core fight is whether reducing cyber capability in training is responsible deployment or a convenient excuse. Supporters argue that removing or downweighting offensive expertise produces a genuinely safer open model than fragile behavioral guardrails; skeptics argue the lab is asking users to trust an unprovable claim of stronger withheld capability, likely shaped by regulation and state interests.
Coverage & discussions
Hacker News · kimi.ai
Community discussions
Community Sentiment
Positives
- Published results and observed outputs make K3.1 look plausibly frontier-level, with commenters calling the jump striking enough to demand closer independent testing.
- A lower API price than K3 despite a larger reported model makes high-end reasoning feel less like an expensive big-lab privilege.
- Training away offensive cyber expertise could make open weights structurally harder to repurpose for attacks than models protected only by refusal layers.
- Existing K3 users value its distinctive strengths and personality, saying it sometimes catches things that other leading models miss.
Concerns
- The headline benchmark claims remain hard to trust when the full results page fails to load for some browsers and practical domain-by-domain comparisons are scarce.
- Restricting cyber knowledge frustrates defensive users who want capable models for legitimate security testing and find heavily guarded alternatives unhelpful.
- Claims of an even stronger internal cyber model invite skepticism because outsiders cannot verify whether capability was responsibly limited or merely withheld.
- Reported paid-plan quotas can cut off an ordinary session within an hour, undermining the model's usefulness for sustained agentic or professional work.
