Kimi K3.1 Pro drops 2.8T open weights on Hugging Face

Downloading a 70B open-source model feels pretty normal these days, but downloading a single model repository weighing in at 1.56TB is a whole different reality check.

Moonshot AI just delivered on its promise by dropping the full weights for Kimi K3.1 Pro directly onto Hugging Face. It is the flagship entry in the new K3.1 series and instantly ranks among the largest multimodal models openly accessible to developers.

Training an artifact of this magnitude easily burns through millions of dollars in compute. Putting the uncompressed weights out in the open has turned heads across both the research community and enterprise infrastructure teams.

Massive scale under the hood

Kimi K3.1 Pro is built on a Mixture-of-Experts architecture containing 2.8 trillion total parameters.

Despite that astronomical footprint, the model routes tokens dynamically, activating only 104 billion parameters during any single inference step to balance compute demand.

It ships with a massive 1-million-token context window. To keep memory and compute manageable at that length, Moonshot AI swapped full-attention layers with 24 Gated DeepSeek Sparse Attention layers across its 93-layer depth.

Multimodal features are baked directly into the core using the 401M-parameter MoonViT-V2 vision encoder, backed by quantization-aware training using MXFP4 weights and MXFP8 activations.

On benchmarks spanning SWE Marathon, Terminal-Bench, and real-world agent evaluations, Kimi K3.1 Pro consistently trades blows with frontier closed-source models like GPT-6 Astra and Claude Opus 5.5.

Enterprise hardware required

While the model is technically public, running it yourself is not for the faint of heart.

At 1.56TB for the complete weights, consumer GPUs and home setups are out of the question. You will need a serious cluster of enterprise accelerators just to get the shards loaded into VRAM.

Moonshot AI has already provided official serving recipes for vLLM and SGLang alongside ready-to-run Docker images, ensuring that engineering teams with proper clusters can spin up inference right away.

Most individual builders will naturally stick to cloud providers or HuggingChat interfaces for everyday experiments.

Why it matters

For enterprise teams and specialized developers, this release is a massive win. You can now host frontier-grade coding agents, complex document workflows, and multimodal analysis entirely behind your own firewall without sending sensitive IP to commercial API endpoints.

For the broader AI ecosystem, it moves the goalposts for open weights. Frontier-tier reasoning used to be an exclusive perk of private labs behind closed APIs, but a 2.8T MoE release proves open architectures can match proprietary giants blow for blow.

We think Moonshot AI played this release brilliantly. By freely sharing an asset that cost fortunes to train, the team locks in immense community goodwill while letting enterprise infrastructure demands drive high-value downstream adoption.