News

Releases & Models

CONFIRMED: Moonshot launches Kimi K3.1 Pro, a 1M-context model that nears Claude Opus 5.5 on coding

Moonshot AI launched Kimi K3.1 Pro and a Highspeed version on Oct 23, already available in the app, Kimi Code and the API, with a 1-million-token context. Benchmarks are the company's own and open weights are only promised by November 4.

Ilustração editorial de uma lua crescente feita de circuitos sobre um data center, com engrenagem 3D e gráfico financeiro holográficos, representando o Kimi K3.1 Pro da Moonshot AI

Moonshot launches Kimi K3.1 Pro and a Highspeed version

Moonshot AI announced Kimi K3.1 Pro on October 23, 2026, the first model in the K3.1 series and, according to the company, the most capable in its 3-trillion-parameter-class open-weight line, with a 1-million-token context window. Alongside it came Kimi K3.1 Pro Highspeed, which Moonshot says is 2 to 3 times faster. Both are already available on Kimi.ai and the apps, Kimi Work, Kimi Code and the API, under the identifiers kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The full weights are not out yet: the company promises to publish them by November 4.

What changes from K3

K3.1 Pro is the same K3 with much more reinforcement post-training spread across hundreds of domains. Moonshot highlights three fronts: long-horizon coding, 3D creation and financial research. On Terminal-Bench 4.0 the company reports 65.7%, ahead of GPT-6 Astra's 58.2% and just behind Claude Opus 5.5's 66.4%. On BenchCAD, which measures 3D reconstruction through CAD code, the model scores 92.7% with tools, against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra. In finance, Moonshot claims the top result on OfficeQA Pro v2 at 64.4 and says the hallucination rate on its hardest internal test fell from 23.4% on K3 to 8.3%.

A deliberate cut in cyberattack capability

The most unusual part of the announcement is the security section. Moonshot says it identified and removed cyberattack-related data from two training stages before the open-weight release. On its internal vulnerability exploitation test, the model without that cut solved 90.5% of tasks; the final K3.1 Pro solves 81.0%. The company says it does not want to release weights that extend the frontier of offensive capability and that this level remains below closed models that have been available for at least three months.

Why it matters for businesses

For anyone using AI at work, the message is price and control. A model approaching closed ones on coding and financial spreadsheet tasks, with weights that can run on your own infrastructure, puts pressure on the cost of relying only on American APIs. The API already offers zero data retention, and in Kimi Business access to K3.1 Pro is off by default until an administrator enables it. Teams handling sensitive data gain another serious option, provided the weights license, not yet published, allows the intended commercial use.

MaxAssistant's take

The launch is real and live, but every number comes from Moonshot. Several coding results were measured with the Claude Code harness, and comparisons with closed models mix in-house tests with third-party scores. K3 also posted a 51% hallucination rate in an independent test, far from what the company advertised. Until independent evaluations arrive, K3.1 Pro is a strong candidate for testing on code and finance tasks, not a declared winner. And open weight only truly applies once the files appear, promised by November 4.

Sources

Moonshot AI: official Kimi K3.1 Pro tech blog; Kimi API Platform.