Moonshot AI Drops Free Weights for 2.8T-Parameter Kimi K3.1 Pro

Moonshot AI just gave away the crown jewels. On November 4, the company published the complete weights for Kimi K3.1 Pro on Hugging Face at the repository moonshotai/Kimi-K3.1-Pro — no waitlist, no approval queue. The download weighs 1.56 terabytes across 96 safetensors shards, and the model packs 2.8 trillion total parameters.

Only about 104 billion of those parameters fire on any given token, thanks to a mixture-of-experts design. That gap is the whole ballgame: a 2.8-trillion-parameter label sounds like a data-centre-only monster, but the active count is what sets inference cost, and 104 billion is a number plenty of serious labs can stomach. Moonshot first put K3.1 Pro behind an API on October 23, where it scored 55 on the Artificial Analysis intelligence index — three points behind Claude Opus 5.5 at 58 and ahead of most open models on the board.

According to FourWeekMBA's analysis of the release, K3.1 Pro actually leads Opus 5.5 outright on 16 of the 37 individual tests making up that composite, even while losing the average. That's a Chinese open-weight model taking nearly half the rounds against Anthropic's flagship. The pricing half of the story is uglier for incumbents: Moonshot has been selling API access at roughly half what Opus 5.5 costs, and now buyers can skip the meter entirely and run it themselves.

Provided they have the gear. Running a 1.56-terabyte model at usable speed calls for something like eight Nvidia B300 GPUs and 2.3 terabytes of combined GPU memory — a tab bigger than most startups' entire annual cloud budget. This isn't a weekend hobbyist project. It's built for hyperscalers, well-funded labs and inference providers angling to cut out the API markup.

The weights don't ship under a standard open-source licence. Moonshot's own Kimi K3.1-Pro License allows use, modification, fine-tuning and redistribution, but any company offering the model as a hosted API service with more than $20 million in trailing 12-month revenue must sign a separate commercial agreement. Researchers and small startups are untouched; the Amazons, Microsofts and Togethers of the world are the target.

K3.1 Pro handles a 1,048,576-token context window — matching the million-token mark Moonshot has used across recent Kimi releases — plus multimodal input including images. Moonshot's release notes pitch it for coding, agentic workflows and financial research, the same turf OpenAI, Anthropic and Google have been fighting over. The company also says it won't release open weights that push the absolute frontier of cyber capability, keeping K3.1 Pro's offensive security skills behind proprietary frontier models that have been public for at least three months — a deliberate throttle aimed at staying ahead of regulators rather than an oversight from the company, as stated in the report. It lands as the fourth major Chinese open-weight release at genuine frontier scale in recent months, following Tencent's Hunyuan, Meituan's LongCat and Alibaba's Qwen3 family. K3.1 Pro is the largest yet by total parameter count, eclipsing DeepSeek's V4 Pro at 1.6 trillion parameters. Washington has spent the past year tightening chip export controls aimed at slowing exactly this — the weights keep going out the door anyway.