Intel / article

Kimi K3.1 Pro ships with 1M context and open weights due Nov 4

Source: kimi.ai
Author: Kimi_Moonshot
Date: Published Oct 23, 2026

From the Daily Brief

Moonshot released Kimi K3.1 Pro, a 2.8T-parameter mixture-of-experts model with about 104B active parameters and a 1M-token context window, through its API at $2 input and $10 output per million tokens. A Highspeed tier runs two to three times faster, and Moonshot says the weights will be published by November 4. On Artificial Analysis's Intelligence Index it scores 55, fourth overall, and its Terminal-Bench 4.0 result rose from 13% to 67% against K3, with cost per task near $1.98. Simon Willison notes the per-token price sits at about half of Opus 5.5. Nathan Lambert calls it the strongest open-weight benchmark showing yet, with two qualifiers: the weights have not shipped, and the model spends more tokens per answer, so cheaper tokens do not translate one for one into cheaper tasks. Artificial Analysis also puts it first among openly available models on its Cyber Index at 64%, against 69% for trusted-access GPT-6 Sol, while Moonshot says it removed offensive cyber data during training. The headline benchmarks are vendor-reported or single-lab results, and the open-weight claim stays a promise until the November release.

Read the 2026-10-24 Brief →

Context

Kimi K3.1 Pro is Moonshot AI's new flagship, and what sets the release apart is that Moonshot deliberately held back its hacking ability before publishing the weights. According to Moonshot's announcement, the model is a version of Kimi K3 trained further with substantially more post-training compute. It reads up to a million tokens of context and, per the launch post, has 2.8 trillion parameters, of which 104 billion are active for any given token. It is available now in Kimi's apps, Kimi Code and the Kimi API, alongside a Highspeed variant Moonshot says runs two to three times faster. The full weights, the files that let anyone download and run the model on their own hardware, are promised by November 4.

Moonshot says the model matches or approaches the strongest proprietary models on coding, agent and knowledge benchmarks. Its headline example is Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks: K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%. Moonshot ran its own model inside the Claude Code and took the other scores from published reports or the official leaderboard, so these are company-reported comparisons rather than one independent run.

The cyber section explains the release decision. Moonshot says vulnerability-research kept rising as post-training scaled, and it traced that to cyber material mixed into the general training corpus and cyber tasks in its reinforcement-learning environments. For the open-weight release it removed that data from both stages while keeping general coding tasks. On xdev-bench-hard, Moonshot's in-house test in which an agent must write a working exploit for a recently disclosed vulnerability in software such as the Linux kernel or Node.js, a checkpoint trained without those removals solved 90.5% of problems, while the released model solved 81.0%. Moonshot says that still leaves K3.1 Pro behind proprietary models that have been public for at least three months, and states its policy as not releasing open weights that extend the absolute frontier of cyber capability while raising what open models can do over time. Because downloaded weights run on infrastructure the user controls, the restraint has to live in what the model learned rather than in a provider's filter.

The evidence for that restraint is Moonshot's own benchmark, and the weights are not yet out. A separate measurement covers defensive work rather than exploitation: Artificial Analysis ranks K3.1 Pro first among openly available models on its agentic cyber defense index.

Key takeaways · AI-distilled

Moonshot says K3.1 Pro is Kimi K3 with substantially more post-training compute and joint RL across hundreds of domains. It reports 65.7% on Terminal-Bench 4.0 with the Claude Code harness, versus 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5.
In an internal test, K3.1 Pro built a human walking simulator, refining the controller over 167 missions and 420 reviews until the simulated human walked for 11.2 seconds instead of collapsing. When reward shaping stalled on arm motion, it found and fixed 36 moment-arm errors.
On Moonshot's internal Kimi Hallu Bench Hard, the claim-level rate fell from 23.4% for K3 to 8.3%. Moonshot cautions the set is a deliberately adversarial stress test, so these rates do not represent everyday error rates.
Moonshot traced rising vulnerability-research skill to cyber content inside its general mid-training corpus and RL environments, and removed that data from both stages for the open-weight release while keeping the general coding tasks.
On Moonshot's in-house xdev-bench-hard, a checkpoint without that data removal solved 90.5% of exploit tasks at pass@3 versus 81.0% for the released model. Moonshot says K3.1 Pro's cyber capability still trails frontier proprietary models public for at least three months.

Terms in this piece · Glossary

agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.

Read the source kimi.ai