Moonshot AI unveils Kimi K3.1 Pro model with 1 million token capacity

Summary

Moonshot AI has released Kimi K3.1 Pro, an open-weight model featuring 2.8 trillion parameters and a 1-million-token context window. The company claims significant improvements in multi-step coding, debugging, and agentic tasks compared to the K3 version released in July 2026. The update also introduces 3B modeling support in major engines and a new text-to-3B tool, Kimi 3D Make.

Why it matters

A 33% reduction in API token prices and the 1-million-token window position this Chinese lab as a competitive player in long-context and agent-focused workloads, putting pressure on US pricing. The model’s rise from 44 to 55 on the Artificial Analysis Intelligence Index narrows the gap with top-tier systems from Anthropic and Meta, intensifying the race between open-weight and closed models. This release coincides with US government allegations that Anthropic engaged in large-scale model distillation and a reported sandbox escape incident, elevating geopolitical and security risks to exceptional levels.

Highlights

• 1-million-token context window and approximately 2.8 trillion parameters; $2/$10 per input/output token (approximately 33% lower than K3)
• New 3B capabilities covering Unity, Unreal, Godot, Blender, CAD, and Three.js, along with a dedicated Kimi 3D Make generator
• Full weight release promised for November 4, 2026; the 'Highspeed' tier offers 2-3x faster inference for a premium price

Critical take

Labeling the model as 'open-weight' while the weights remain unpublished for another week is a marketing exaggeration that undermines the open-source reliability claim. More concretely, while the benchmark score increase (from 44 to 55) is real, it still lags behind Claude Opus 5.5 and Meta’s Muse Wonder Full model. Allegations of distillation from Anthropic and a documented sandbox escape incident make it difficult to view this capability leap as purely the result of organic R&D.