THE SCOREBOARD FOR WHAT'S NEXTAI + QUANTUM, SCORED. COMPARED. EXPLAINED.
AIQUANTUMSCORE.
← AI & Quantum News

AI / NEWS BRIEF

Moonshot releases Kimi K3.1 Pro's full model weights

The downloadable 2.8-trillion-parameter checkpoint fulfills Moonshot's October promise, but its custom licence and roughly 1.56-terabyte footprint make the release open-weight rather than frictionless.

By AIQuantumScore · · Event 2026-11-04

The promised checkpoint is now downloadable

Moonshot AI has published Kimi K3.1 Pro's model files on Hugging Face, fulfilling the October launch promise to release full weights by November 4. The public, ungated repository describes an open-weight, natively multimodal agent model and provides deployment instructions for vLLM, SGLang and TokenSpeed. Independent release tracking confirmed that the files became publicly available on November 4.

Large even in its native low-precision format

Moonshot's model card lists 2.8 trillion total parameters, 104 billion activated per token and a 1,048,576-token context window. The released checkpoint uses MXFP4 weights and MXFP8 activations produced through quantization-aware training. ArtificialWatch counted 96 safetensors shards totaling about 1.56 terabytes, so running the release still requires substantial multi-GPU infrastructure despite the mixture-of-experts design.

Open-weight does not erase the licence or evaluation caveats

OUR ANALYSIS

The repository uses Moonshot's custom Kimi K3.1 Pro licence rather than a standard open-source software licence. The release enables independent hosting and testing, which is materially different from API-only access, but it does not independently validate Moonshot's benchmark claims. The practical significance is access: researchers and infrastructure providers can now inspect and serve the actual checkpoint, while users should review the licence and wait for reproducible evaluations of the downloaded files before treating vendor performance tables as settled.

More to follow