Kimi K3.1 Pro weights are out. Self-hosting starts at 16 GPUs

Moonshot put the full Kimi K3.1 Pro weights on Hugging Face on November 4, days before Anthropic and Google spent the same week announcing closed models with closed math. The download is free. Running it is not. The checkpoint is 1.56 TB across 96 safetensors shards, and we measured the files directly rather than trusting the launch post: 1,561,293,160,568 bytes, every one of them has to sit in GPU memory, because the router picks a different 16 experts out of 896 for every token. Moonshot's own API charges $2 per million input tokens and $10 per million output. The question this post works through: when does the download beat the API?

The model card gives the shape: 2.8 trillion total parameters, 104 billion active per token, a 1M-token context window, text and image input. The weights ship as MXFP4, already 4-bit, so there is no smaller checkpoint coming and re-quantizing below 4 bits costs quality. Yotta Labs walked the GPU ladder: a single 192 GB card holds about an eighth of the weights, a full H100 node holds less than half, and the smallest shape that boots at all is 16x 141 GB across two H200 nodes with roughly 2.26 TB of aggregate memory. That configuration leaves about 700 GB for KV cache, enough for evaluation and moderate traffic at reduced context. Serving the full 1M window to real users means the 32-card tier or above.

Now the money. A 16x H200 cluster rents for roughly $40 per hour on current market rates, and that node spends most of its capacity just holding the model. Serve a modest 5 tokens per second of steady output and the node produced 18,000 tokens in that hour: roughly $2,200 per million output tokens, more than 200 times Moonshot's API rate, before you have paid for a single prefix cache hit. Moonshot's API sells the same output at $10 per million with cached input at $0.20. The API is not the expensive option here. The expensive option is the cluster you rented and let idle.

So who is the download for? Three groups, and the license draws the same lines. The Kimi K3.1 Pro License is a modified MIT, and the raw text on Hugging Face sets two thresholds: a business reselling model access needs a separate agreement with Moonshot once its aggregate revenue passes $20 million over any rolling 12 months, and any product above 100 million monthly active users must display "Kimi K3.1-Pro" in its interface. Below those lines, commercial use is free and attribution is the only obligation. Fine-tuning on private data, keeping prompts off other companies' servers, building on an air-gapped cluster: those are real reasons to eat the hardware bill. Note the pattern, though, because it repeats: Meta's Glimmer weights in August, DeepSeek's V4.1 flash release in October, now Kimi. The open-weights frontier keeps arriving with an invoice attached, 1.56 TB of it this time, and the free part stops at the download page.

The benchmark claims stay unreplicated. Moonshot's own tables put K3.1 Pro at 65.7 percent on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2 and a point behind Claude Opus 5.5, with the strongest agentic rows on AutomationBench-AA at 73.3. Every figure comes from Moonshot's own evaluation stack with reasoning effort set to max, and the same caveat applies to every lab's self-reported numbers this month. For comparison, Gemini 4 Carbon went generally available on November 10 at $2 per million input, the same number Kimi charges. Whether that 73.3 survives independent testing is exactly the kind of question a replicated benchmark answers, and Uteke's benchmark posts make our own position on vendor-reported numbers plain.

Two dates matter if you are pricing work this month. Carbon's $2 rate is introductory: after November 30 it rises to $4 per million in and $20 out, while Kimi's API holds the cheap end of the frontier at $2 in and $10 out. The weights cost nothing to copy and about $40 an hour to wake up. Price your tokens accordingly.

Sources: moonshotai/Kimi-K3.1-Pro weights and model card (Hugging Face, Nov 4), the license text quoted above, Kimi K3.1 Pro tech blog (Moonshot AI, Oct 23), Kimi API pricing, GPU ladder analysis (Yotta Labs, Nov 5), Gemini 4 Carbon availability (Google, Nov 10) with the Carbon model card, Fable 5.2 launch coverage (CNBC, Nov 9), and the Hacker News thread.