Moonshot AI Releases Kimi K3.1 Pro: A 2.8T-Parameter MoE Model with 1M-Token Context
Moonshot AI has released Kimi K3.1 Pro, the first model in its K3.1 series. It is a 3T-class open-weight model with a 1-million-token context window, positioned against the strongest proprietary models on coding, agentic, and knowledge benchmarks. For developers, it offers cheaper API pricing than K3, though open weights are not yet available. Those weights are expected to be released in November 2026.
Key Takeaways
- Scale: 2.8T total parameters with 104B active parameters (Mixture of Experts), plus a 1M-token context window.
- Deployment: Available today via the Kimi API. For K3, Moonshot recommends supernodes of 64+ accelerators using MXFP4 weights and MXFP8 activations—this is not a single-GPU model.
- Performance: Close to the proprietary frontier on terminal coding tasks, but behind it on scientific terminal tasks.
- Strongest result: SWE-Marathon score of 51.7, ahead of every model in Kimi's chart (self-reported).
- Weakest result: Terminal-Bench Science 0.1 score of 38.1, versus 68.1 for GPT-6 Astra in Kimi's chart.
What Is Kimi K3.1 Pro?
Kimi K3.1 Pro is a sparse Mixture-of-Experts (MoE) model that activates only a fraction of its parameters per token. With 2.8T total parameters and 104B active parameters, it targets efficient inference at frontier scale—a pattern now standard among 2026 flagship open-weight releases. The 1M-token context window places it in the same tier as leading long-context proprietary models and supports agentic and repository-scale coding workflows.
Moonshot frames the K3.1 series as a competitor to top proprietary systems on coding, agentic, and knowledge benchmarks. Developers can access the model through the Kimi API at lower pricing than K3, while awaiting the open-weight release in November 2026.
via MarkTechPost