Tech
Moonshot opens Kimi K3.1 Pro weights: 2.8T parameters, 104B active, 1M context.
Moonshot AI · Hugging Face · 2026-11-04 · Source ↗
Moonshot opens Kimi K3.1 Pro weights: 2.8T parameters, 104B active, 1M context.
Moonshot AI · Hugging Face · 2026-11-04 · Source ↗
Kimi K3.1 Pro is the first weights release at this scale from Moonshot since the K3 generation opened in September. The model carries 2.8 trillion total parameters and 104 billion active parameters per token, distributed across 93 layers (one dense) with 896 experts and 16 routed per token. The architecture is multimodal from the first token, accepting both text and image inputs, and the context runs to 1,048,576 tokens.
The quantization choice
Moonshot shipped the weights in MXFP4 with MXFP8 activations under a quantization-aware training regime, which is a deliberate choice against the more common bfloat16 open release. The practical consequence is that the weights file is roughly a quarter the size of a bf16 version at the same scale, and inference can run on narrower hardware without requiring post-training calibration. The tradeoff is that fine-tuning tooling must support MXFP4 natively or re-upcast; most open frameworks support one but not both.
Self-reported benchmarks
The model card reports Terminal-Bench 4.0 at 65.7, SWE Marathon v1.1 at 51.7, BrowseComp at 92.1, GPQA-Diamond at 94.2, HLE-Full with tools at 70.0, OSWorld V2.1 at 79.6 and an internal Kimi Code Bench 2.0 at 77.1, benchmarked against Claude Opus 5.5, GPT-6 Astra and DeepSeek-V4.1-Flash. The scores land at or near parity with Opus 5.5 on several agentic benchmarks and sit above DeepSeek-V4.1-Flash across the board. All but GPQA and BrowseComp are reported by Moonshot without third-party replication.
"MXFP4 weights and MXFP8 activations, trained with quantization-aware training." — Kimi K3.1 Pro model card, 2026-11-04
License and availability
The release is under a custom "Kimi K3.1 Pro License" rather than Apache 2.0 or MIT, which is consistent with Moonshot's prior releases. The model card does not state explicit commercial-use terms; readers should consult the LICENSE file on the Hugging Face repository. No arXiv paper accompanied the release.