China just dropped a new frontier model... and nobody can explain how

In progressFireship (The Code Report)video2026-10-26read planted ai llmssecurity

Synopsis — AI-drafted from Dan's notes

A six-minute Code Report, uploaded on 26 October 2026, about Kimi K3.1 Pro from Moonshot AI. An anonymous model called Kibitz had turned up on Arena the week before, and testers guessed it was Kimi (Testing Catalog). Moonshot released K3.1 Pro through its apps and API on 23 October and promised the open weights by 4 November. Fireship calls it a 3-trillion-parameter model; Moonshot describes it as 3T-class, and coverage gives 2.8 trillion parameters in total, 104 billion active.

His hook is distillation. In February 2026 Anthropic accused Moonshot, DeepSeek and MiniMax of distilling Claude through thousands of fraudulent accounts, and he says Anthropic now claims to block 90 to 99% of such attempts. If the big labs are right that Chinese models catch up by copying, he argues, a model this good should not be possible now; along the way he defends distillation, half in jest, as fair payback for training on everyone’s data. On Moonshot’s own figures K3.1 Pro scores 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5, and leads SWE Marathon (51.7 to Opus’s 46.3). He says Artificial Analysis independently measured 67%, and warns that the harness changes these numbers a lot.

The caveat he dwells on is cyber. Moonshot stripped cyber material from mid-training and its RL tasks: an unmitigated checkpoint solved 90.5% of its exploit benchmark, the released model 81%. He reads this as self-protection, since the Department of Homeland Security had just ordered DeepSeek products out of federal civilian agencies (published around 19–20 October). He closes on his joke benchmark, a one-prompt 3D horse in three.js: Kimi’s horse is the prettiest but animates by warping the mesh, where Claude built a rigged skeleton, so Kimi scores 4.20, behind Opus (6.39) and GPT-6 Astra (5.21).

Connections