Kimi K3.1 Pro open weights are live, with a $20M MaaS clause

Moonshot AI's promised open-weights release for Kimi K3.1 Pro is no longer a promise. The full weights of the 2.8-trillion-parameter mixture-of-experts model are verified live and ungated on Hugging Face as of this morning, which moves one of the strongest API-only models of the cycle onto self-hostable hardware.

Key facts

• The official moonshotai/Kimi-K3.1-Pro repo is ungated, with 112 files including 96 safetensors shards, ~11.7k downloads and 944 likes at DeAI's check today
• The model card lists 2.8T total parameters, 104B activated, 93 layers, and 896 experts (16 selected plus 2 shared) per token
• Context window is 1,048,576 tokens, with weights in MXFP4 and activations in MXFP8 under quantization-aware training
• The license requires MaaS operators above $20M aggregate annual revenue (12 months, including affiliates) to sign a separate agreement with Moonshot before commercial use
• OpenRouter already lists moonshotai/kimi-k3.1-pro at $2.00 per million input tokens and $10.00 per million output tokens, matching the Kimi API's launch pricing
• SGLang day-0 serving support for K3.1 Pro is in flight, alongside a separate open PR for ERNIE 5.5 Flash

What happened

On October 23, Moonshot AI launched Kimi K3.1 Pro through its API and committed to releasing the full weights by November 4. The deadline came and went with no official repo, which we reported in Wednesday's deadline check. That cliffhanger is now resolved: the moonshotai Hugging Face org carries the Kimi-K3.1-Pro repository, and DeAI verified it directly this morning — 96 safetensors shards among 112 files, no access gate, roughly 11.7k downloads already on the counter.

One timing caveat belongs up front. The repo's Hugging Face metadata shows a creation date of November 2, yet our November 4 org check found no K3.1 Pro listing. The likeliest explanation is that the repo existed privately before it was publicly listed. Because the listing moment is what we can verify, the accurate statement is this: the weights are confirmed live as of November 6, and we can't say exactly when they first became public.

The specs on the model card match everything Moonshot promised at launch. Kimi K3.1 Pro is a 2.8T-parameter mixture-of-experts model with 104 billion parameters activated per token, built from 93 layers and 896 experts (16 selected plus 2 shared) with a gated dynamic-sparse attention mechanism replacing gated MLA for long context. The context window is 1,048,576 tokens. Weights ship in MXFP4 with MXFP8 activations, meaning the model was trained quantization-aware rather than quantized after the fact, and a separate 401M-parameter MoonViT-V2 vision encoder gives it native image input.

The license is the detail builders should read before planning anything. The Kimi K3.1 Pro License lets most operators use the weights commercially without negotiation, but it draws two bright lines. Model-as-a-service operators with more than $20 million in aggregate revenue over the trailing 12 months, counting affiliates, must sign a separate agreement with Moonshot before commercial use. Products with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3.1-Pro" branding. Internal use and Moonshot-certified partners are exempt from both.

Why it matters

Self-hosting a frontier-class model stopped being theoretical with DeepSeek and Kimi K3; K3.1 Pro extends that roster at the very top. A 3T-class MoE with a 1M-token context in a self-hostable download changes the build-versus-buy math for anyone running long-document or agent workloads where per-token API spend compounds. The activation count is the number that matters for capacity planning: 104B active parameters is the serving cost, not 2.8T, though the full parameter set still has to sit in memory.

The license clauses decide who can sell access to it. Below the $20M threshold, a provider can serve K3.1 Pro without talking to Moonshot. Above it, serving requires a signed agreement, and marketplace-scale deployments carry a branding obligation. If you operate inference infrastructure or are modeling provider listings, read the license text on the page rather than relying on summaries, including this one.

For the open-weights ecosystem, this is the second time in a week a top-tier lab has converted a weights commitment into a delivery. Thinking Machines Lab said yesterday that Inkling 2 Small will ship full weights under Apache 2.0 next week after external red-teaming, and Mistral's Large 4 open-weights drop earlier in the cycle followed a similar pattern. Commitments are becoming the norm; the interesting variance is now in the license, not the download.

Background

Kimi K3.1 Pro launched October 23 as an API-only model at $2.00/$10.00 per million input/output tokens, scoring 55 on Artificial Analysis' independent Intelligence Index at the time — third among tracked models, behind Anthropic's Claude Opus 5.5. That index score came from the API model, not the released weights. The Terminal-Bench 4.0 result and agent rankings Moonshot circulated at launch are self-reports, and independent benchmarks on the downloadable weights have not surfaced yet. As we noted in the launch coverage, the open-weights release was the whole point of the launch: if the released model reproduces the API's scores, it becomes the highest-scoring open-weight model anyone can actually download.

The download also lands amid a broader push by Chinese labs to get weights into serving stacks fast. SGLang pull requests for K3.1 Pro nightly support and docs landed November 4 and 5, and an open PR adds ERNIE 5.5 Flash from Baidu, whose own base-model weights appeared on Hugging Face in late October. Day-0 engine support shortens the gap between a weights drop and a servable deployment to days, which is roughly how long it took for OpenRouter to list the model at the API's launch price of $2/$10 per million tokens. Our open-weights hub tracks how license classes — MIT, Apache 2.0, community licenses, and now this modified-MaaS structure — shape who can deploy what.

What's next

The near-term milestones are all verification work. Independent benchmark runs on the released MXFP4 weights should appear within days, and until they do, the 55-point Intelligence Index score belongs to the API model, not the download. Engine support will decide how quickly the model reaches production quality on self-hosted stacks: the SGLang PRs are in flight, and vLLM support tends to follow the same cadence. And the Inkling 2 Small Apache 2.0 commitment, due next week, will test whether the weights-commitment pattern holds for a US lab the way it just did for Moonshot. Watch the license negotiations, too: the first MaaS provider to sign above the $20M line tells you where Moonshot plans to draw the commercial boundary.