Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context
MarkTechPost Asif Razzaq ● Covered by 2 sources
Moonshot AI launched Kimi K3.1 Pro, a 2.8T MoE model with a 1M-token context. It’s already on the API, but the weights won’t be public until later.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Moonshot AI has launched Kimi K3.1 Pro, the first model in its K3.1 line, and it arrives with an eye-catching number attached: a 1-million-token context window. The company is framing it as a 3T-class open-weight model, though the weights are not public yet. For now, developers can only use it through Moonshot’s own services.
The model itself is a mixture-of-experts system with 2.8 trillion total parameters and 104 billion active parameters. Moonshot says it is a post-trained version of Kimi K3, built with much more post-training compute and joint reinforcement learning across hundreds of downstream domains. Under the hood, the company also says it improved RL infrastructure, long-context inference, sandboxing and rollout scheduling.
On benchmarks, Moonshot is making a direct push at the proprietary front. In its own chart, Kimi K3.1 Pro scores 65.7 on Terminal-Bench 4.0, just behind Claude Opus 5.5 at 66.4, and leads SWE-Marathon with 51.7. But the same chart shows a clear weakness in science-heavy terminal work: 38.1 on Terminal-Bench Science 0.1, far below GPT-6 Astra’s 68.1 and Opus 5.5’s 63.3.
There are other bright spots. Moonshot reports 92.7 on BenchCAD with tools and 72.8 without them, plus a 64.4 on OfficeQA Pro v2. The company also says claim-level hallucinations on its internal Hallu Bench Hard test dropped from 23.4% in K3 to 8.3% in K3.1 Pro. And on cyber risk, Moonshot says it stripped cyber-related data from mid-training and reinforcement learning for the open-weight release, while the shipped model still trails frontier proprietary systems on offensive tasks.
Kimi K3.1 Pro is live now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, with OpenAI-compatible endpoints. Pricing is lower than K3: $2.00 input and $10.00 output per million tokens, with cache hits at $0.20. Moonshot also offers a Highspeed version that it says runs 2 to 3 times faster, but the ordinary model is not a single-GPU toy. Moonshot recommends supernodes of 64 or more accelerators for K3-class serving, which is the sort of detail that quietly tells you this model is not meant for a laptop on a café table.
My take — AI-written commentary, not fact-checked reporting
Moonshot is doing the sensible thing here: ship access first, promise the open weights later, and let everyone else argue about purity. The bigger story is that frontier model economics now look like a pricing war wrapped around a hardware problem. Open-weight in spirit, closed in practice — the usual modern compromise, with a nicer API bill.
Read more about this at: MarkTechPost
