Skip to content
Neural Daily

ModelsNews

Kimi K3.1 Pro: Moonshot's open-weight model nears Opus

🇭🇺 Olvasd el magyarul →

Moonshot AI released Kimi K3.1 Pro, a 2.8-trillion-parameter MoE model with a 1-million-token context window. It ranks 4th on the Artificial Analysis index, and the weights are due by November 4.

By: Mikola Máté Patrik · · 3 min read

Kimi K3.1 cover: stylized K3.1 lettering over misty mountain peaks, with the tagline From Ideas to Finished Work
Image: Moonshot AI · Press kit · Official announcement social image (og:image) · Source

What launched?

On October 23, China's Moonshot AI released Kimi K3.1 Pro (Kimi), the first model in its K3.1 series and, by the company's account, its most capable 3T-class open-weight model yet. Alongside it comes Kimi K3.1 Pro Highspeed, which offers the same intelligence at 2-3x the speed.

  • Architecture: MoE with 2.8 trillion total and 104 billion active parameters, 1-million-token context (per Artificial Analysis).
  • A post-trained version of K3 with much more post-training compute, using large-scale joint RL across hundreds of domains.
  • Available on Kimi.ai, in Kimi Work and Kimi Code, and via the API (kimi-k3.1-pro and kimi-k3.1-pro-highspeed).
  • Pricing: $2 / $10 per million input / output tokens, $0.20 for cache hits, 33% lower than K3 on every rate.
  • Full weights will be released by November 4. The license has not been announced.

What the benchmarks show

On Moonshot's own numbers, it scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). On BenchCAD (3D reconstruction through CAD code generation) it reaches 92.7% with tools versus 96.2% for Opus 5.5, and it leads the table on the OfficeQA Pro v2 financial document benchmark with 64.4.

Moonshot AI benchmark overview comparing Kimi K3.1 Pro with Opus 5.5, GPT-6 Astra, Fable 5.1, DeepSeek-V4.1-Flash-1017 and GLM-5.3
Moonshot AI · Press kit · Official benchmark chart from the Kimi K3.1 Pro blog · Source

Source: Moonshot AI. The company's own benchmark overview, run at maximum reasoning effort.

Independent tester Artificial Analysis puts it at 55 on its Intelligence Index, 4th behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56), and 11 points above K3. Once the weights are out, it will be the top open-weight model on the index, narrowing the gap to proprietary models from 12 points to 3. AA also measured 67% on Terminal-Bench 4.0, the highest of any model.

On Moonshot's internal hallucination stress test, the claim-level error rate fell from 23.4% to 8.3%. The test is deliberately hard and does not reflect everyday error rates.

Cyber capability, deliberately capped

For the open-weight release, Moonshot removed cyber-related data from both mid-training and RL. On its internal xdev-bench-hard exploit benchmark, the unfiltered checkpoint solved 90.5% of tasks and the released model 81.0%. The company says this keeps its cyber capability behind proprietary frontier models that have been public for at least three months.

Why it matters

If the weights ship as promised, an agentic and coding model close to Opus will be runnable on your own infrastructure, though only on serious GPU servers: K3's weights were about 1.6 TB. Until then the API is the practical route, at a third less than K3. In Kimi Enterprise, admins must switch it on, since access is off by default.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles