Moonshot AI just raised the bar for open-weight AI. The Chinese lab behind the Kimi assistant launched Kimi K3.1 Pro on October 24, 2026 — the first model in its new K3.1 series — and it is Moonshot’s most capable
open-weight model to date. A faster sibling, Kimi K3.1 Pro Highspeed, launched alongside it: two to three times the speed with identical capabilities, according to the company.
A 3T-class flagship, post-trained on Kimi K3
K3.1 Pro is a post-trained version of Kimi K3, built with what Moonshot calls substantially larger post-training compute. It sits in what the company describes as the 3T-parameter class and carries a 1-million-token context window. To run reinforcement learning across hundreds of downstream domains at this scale, Moonshot says it optimized every layer of the stack — the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for environment quality checks.
The benchmark numbers (company-reported)
On
Terminal-Bench 4.0, which tests coding agents on realistic, resource-constrained terminal tasks, K3.1 Pro scored 65.7 percent — ahead of GPT-6 Astra's 58.2 percent and just behind Claude Opus 5.5's 66.4 percent. On the company's internal long-horizon coding test, Kimi Code Bench 3.0, it scored 61.2 percent against K3's 55.95 percent. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, it hit 92.7 percent with tools — approaching Claude Opus 5.5's 96.2 percent and GPT-6 Astra's 95.9 percent. A 3D Make feature is live on kimi.ai and the Kimi app, letting users preview and edit 3D models online and download them ready for 3D printing.
The less-hallucinating part
Moonshot also reports a sharp drop in hallucinations: on its internal Kimi Hallu Bench Hard stress test, the claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent. In one long-horizon coding demo, K3.1 Pro built a human walking dynamics simulator on top of human-atlas, an open-source anatomical model with 296 skeletal and 402 muscle components — running 167 missions and 420 independent reviews, and correcting 36 errors in shoulder-muscle moment-arm calculations when its first approach stalled. (The usual caveat applies: these are the lab's own numbers; independent runs will tell the fuller story.)
Where you can get it
Both models are rolling out across the Kimi product family: kimi.ai and the mobile apps for iOS, Android, and HarmonyOS; the Kimi Work desktop app (version 3.1.0 or later, for Windows and Apple silicon Macs); Kimi Code via the /model command; and the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available. Membership tiers: k3.1-pro needs Plus and above, the full 1M context needs Pro and above, and the Highspeed variant needs Pro and above. Full model weights drop by November 4, 2026.
The safety move worth praising
Here is the part that deserves more attention. As post-training scaled, Moonshot noticed the model's vulnerability-research performance improving too — traced to cyber-related content embedded in its training. For the open-weight release, the company deliberately removed cyber-related data from both training stages. On its in-house xdev-bench-hard vulnerability-discovery test, the unscrubbed checkpoint solved 90.5 percent at pass@3 while the final model reaches 81.0 percent — and the company states K3.1 Pro's cyber capability stays behind frontier proprietary models that have been public for at least three months.
Moonshot frames two principles behind the release: never put
open weights out that would extend the absolute frontier of cyber capability, while progressively raising the cyber capability that's openly available — so more teams can audit the code they maintain, including legacy systems that rarely get security reviews. A lab measuring how hackable its model is, then training that skill out before handing out the weights, is genuinely good practice. More labs should do this.
Why it matters
Open-weight models are closing in on the frontier fast — and this one does it with hallucination rates cut by nearly two-thirds and a stated safety boundary on cyber capabilities. With Chinese labs shipping aggressively this month, Kimi K3.1 Pro is a reminder that “open” does not have to mean “careless.” Watch the weights release on November 4 — that's when the independent benchmarks start.
Comments
Post a Comment