Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context

Moonshot AI has released Kimi K3.1 Pro, a mixture-of-experts model with 2.8 trillion total and 104 billion active parameters and a one-million-token context window. Moonshot pitches it against the strongest proprietary models on coding, agentic and knowledge work, and it's live now on Kimi's apps and API.

Two things make it newsworthy. API pricing drops to $2 input and $10 output per million tokens, down from K3's $3 and $15. And Moonshot says open weights will follow in November 2026. If a near-frontier coding model becomes downloadable, the economics of agentic software shift.

We'll start with the asterisk the headline deserves: this is not an open-weight model today. It's an API model with a promise. Until the weights land, nobody can self-host or independently probe it, and the open label is a forecast.

On capability, the honest picture is strong but uneven, and all of it self-reported at maximum reasoning effort. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, a hair behind Claude Opus 5.5's 66.4% in Moonshot's chart. It leads SWE-Marathon at 51.7. But on Terminal-Bench Science it scores 38.1 against 68.1 for GPT-6 Astra, a gap of about 30 points. Competitor figures come from Moonshot's own chart and weren't confirmed on rivals' pages, and several wins are on internal benchmarks.

The hallucination claim is the one we most want verified. On an internal stress test, claim-level hallucinations fell from 23.4% to 8.3%. If that holds in the wild, it's meaningful for finance and research users. GPT-6 Astra still posts 5.5% on the same test, so this is progress, not leadership.

The cyber section is where we pay closest attention. Moonshot says it removed cyber-related data from mid-training and RL for the open-weight release, and that an unfiltered checkpoint solved 90.5% of its internal offensive tasks at pass@3 while the shipped model solves 81.0%. We appreciate the disclosure. But read it plainly: the mitigation trimmed a few points, and the model still clears most of that internal test. Moonshot says it stays behind frontier proprietary models on offensive work. Once weights are public, there's no API to pull back. If this pattern scales, capable offensive tooling becomes a download, and the safety story rests on data filtering alone.

We're for open models and cheaper access; they keep the field competitive. We also want independent red-teaming before November, not after.

K3.1 Pro is a heavily post-trained version of Kimi K3, which uses Kimi Delta Attention, Attention Residuals and activates 16 of 896 experts. For K3, Moonshot recommended supernodes of 64 or more accelerators, so this is not a single-GPU model. A Highspeed variant runs two to three times faster at $4 input and $20 output per million tokens. Thinking mode is always on, and Moonshot advises returning the full assistant message in agent harnesses.