AI Quanting (@AIQuanting) on X

What it claims

Moonshot AI released Kimi K3.1 Pro, a post-trained Kimi K3 with a 1-million-token context window, alongside a Highspeed variant that Moonshot says runs 2 to 3 times faster. Both are live on Kimi's apps and API, and Moonshot says full weights will be released by 4 November 2026. For the open-weight release, Moonshot removed cyber content from the mid-training corpus and the RL task mix and reported what the removal cost in exploit capability. The post's author (@AIQuanting) argues the reduction is small.

How it works

Moonshot found cybersecurity content inside its general mid-training corpus and its RL task mix, and stripped it from both stages. Ordinary coding tasks were kept. It then compared a checkpoint with no removal against the shipped K3.1 Pro on xdev-bench-hard, its in-house benchmark. Most xdev-bench-hard tasks rebuild recently disclosed bugs in the Linux kernel, Node.js, JavaScriptCore, Exim and VirGL. The agent must write a working exploit and run it.

Evidence

On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On Moonshot's internal hallucination stress test, the claim-level error rate fell from 23.4% for K3 to 8.3% for K3.1 Pro.

On xdev-bench-hard at pass@3, the no-removal checkpoint solves 90.5% of tasks and the shipped K3.1 Pro solves 81.0%. Moonshot says this shows data removal meaningfully reduces cyber capability. It also says it does not want open weights that extend the absolute frontier of that capability. The author reads the same numbers differently: removal cut 9.5 points, a little over a tenth of the score, so four of five expert exploit tasks still fall to a downloadable model.

Limits

The hallucination test and xdev-bench-hard are both internal Moonshot benchmarks, and the post gives no task counts, harness details or settings for either. The Highspeed variant has no benchmark scores in the post, only the 2 to 3 times speed claim. The post names no weight licence, parameter count or API pricing. Its source is a Kimi blog post linked as kimi.ai/blog/kimi-k3-1, which the thread truncates. The weights were not yet released as of late October 2026.