Kimi K3's Weights Actually Shipped: 1.56TB on Hugging Face, Single Digits Off the Frontier. The 'MIT' License Grew a $20M Tripwire in Transit.

TL;DR

On July 27, exactly as promised, Moonshot AI published the full weights of Kimi K3 to Hugging Face: 96 safetensors files, 1.56TB, natively quantized to MXFP4. The 2.8-trillion-parameter mixture-of-experts model, its million-token context window, and its near-frontier benchmark table were covered here on July 16; what is new today is the artifact itself, and two things you could not read until it existed. Tom's Hardware reads the economics as roughly 2-3x cheaper to run than the frontier models it trails, up to 10x when a query hits the cache. And the license, billed as Kimi K2-style modified MIT, picked up a clause on the way to the repo: operate a model-as-a-service business while your aggregate revenue tops $20 million in any 12 months, from any source, and you need a separate agreement with Moonshot before you can keep selling.

Eleven days from pitch to artifact

Kimi K3 has had a louder July than most models get in a lifetime. The announcement landed July 16 with the benchmark table. On July 23, a researcher pointed 32 K3 agents at Redis and reported an RCE proof of concept in 27 minutes. On July 24, the White House accused Moonshot of distilling K3 from Claude Fable 5 while 200 startups petitioned Washington not to ban Chinese open weights. And on July 27 the weights published anyway.

What is actually in the box

The repo holds 96 safetensors files totaling 1.56TB: MXFP4 weights with MXFP8 activations, quantization-aware from the supervised fine-tuning stage onward. That detail matters. K3 ran the last leg of training wearing the compression suit it ships in, rather than being squeezed into one after the fact, which is why nothing should snap the way it does when a full-precision model gets crushed to 4-bit post hoc.

Per the model card: 93 layers, of which 69 use Kimi Delta Attention and 24 use gated MLA, a 160K vocabulary, a 1,048,576-token context, and native vision (the repo is tagged image-text-to-text). Of the 2.8 trillion parameters, 16 of 896 experts fire per token, so about 104 billion parameters are awake at any moment. Day-one serving support is listed for vLLM, SGLang, and TokenSpeed, and Moonshot's own guidance targets supernode deployments of 64 or more accelerators for the full million-token context.

The license grew a tripwire in transit

When K3 was announced, the working assumption was a rerun of Kimi K2's modified MIT, whose only unusual demand was branding: pass 100 million monthly active users or $20 million in monthly revenue and you must display "Kimi K2" prominently in your product. The actual K3 license keeps that display clause (now reading "Kimi K3") and adds a second, sharper one: if you or your affiliates operate a model-as-a-service business, and your aggregate revenue, from all sources combined, tops $20 million over any consecutive 12 months, you must "enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose." Note what the threshold measures: total company revenue, not just what the model serving brings in. The definition cuts the other way too: the license's "Model as a Service" explicitly excludes embedded end-user features, agent harnesses, and request relays, and internal use and Moonshot's certified inference partners are exempt. Shipping K3 inside your product is fine at any size, branding clause aside. Selling raw K3 access by the token is what arms the wire.

In other words: free in the way a casino buffet is free. Walk in, eat all you want, build whatever you want. But once your company clears $20 million a year and one of the things you sell is raw K3 by the token, the house wants a word. Tom's Hardware waves this off as "few restrictions," which is true for almost everyone reading this, and materially untrue for exactly the class of GPU-rich reseller the release otherwise empowers.

What "2-3x cheaper" actually buys

Tom's Hardware's headline math: K3 closely trails Claude Fable 5 and GPT-5.6 Sol while being "seemingly around 2-3x cheaper to run, up to 10x if a particular query lands in the cache." Moonshot's API prices at $0.30 per million cache-hit input tokens, $3.00 cache-miss, and $15.00 output. The cost structure comes from the architecture: 104B active parameters and 4-bit weights mean far less VRAM and compute per token than the dense frontier bill, and Moonshot claims its Kimi Delta Attention design decodes up to 6.3x faster at million-token context lengths.

The detail worth an eyebrow, spotted by Tom's Hardware: Moonshot's own write-up mentions running some coding tests on Nvidia's H20, the export-control-shaped chip that does not even have native MX floating-point support. The model is efficient enough to be benchmarked on hardware that was designed to be too weak for this.

The capability those dollars buy is not hypothetical. On Moonshot's model card table, K3 posts 88.3 on Terminal-Bench 2.1 against 84.6 for Claude Opus 4.8 and 83.4 for GPT-5.5, 91.2 on BrowseComp against 84.3 and 84.4, and 67.5 on DeepSWE against 59.0 and 67.0. Those are vendor-reported numbers against last-generation frontier models, and Moonshot itself concedes K3 trails Fable 5 and GPT-5.6 Sol on overall intelligence; at launch, Artificial Analysis slotted it fourth on its index at 57, per Glitchwire. Fourth. Behind three closed models, ahead of everything else you are allowed to download.

What changes now that it is a file

When we covered the Redis zero-day run, the point was that on July 27 the thing that found 19 bugs in 90 minutes would stop being an API that can refuse and start being a file that cannot. It is July 27. The weights are seeing thousands of downloads on day one, and as Tom's Hardware puts it, "most anyone with a contemporary rack of AI GPUs can run it and charge for it." Every safety behavior K3 has is now a default, not a boundary.

Two caveats keep their place in the record. Open-weight is not open-source: the training data and process remain, in Tom's Hardware's phrase, Moonshot's special secret sauce, and the White House's distillation accusation is still unresolved. And a homelab reality check: 1.56TB of MXFP4 across 64-plus accelerators is datacenter-class open weights. What lands on your rig will be the community quants and distills, and after two generations of Kimi releases, those tend to arrive fast.

Key Takeaways

• Kimi K3's full weights published to Hugging Face on July 27 as promised: 96 safetensors, 1.56TB, MXFP4 weights with MXFP8 activations, quantization-aware from the SFT stage onward.
• Architecture per the model card: 2.8T total parameters, 104B active (16 of 896 experts), 93 layers mixing Kimi Delta Attention and gated MLA, 1M-token context, native vision, day-one vLLM/SGLang support.
• The license is K2-style modified MIT plus a new clause: operate a model-as-a-service business while aggregate revenue, from any source, tops $20M in any 12 months, and commercial use requires a separate agreement with Moonshot. The K2 license only ever demanded branding.
• Tom's Hardware reads the economics as roughly 2-3x cheaper to run than the frontier models K3 trails, up to 10x on cache hits, with API pricing at $0.30/$3.00 per MTok input (hit/miss) and $15.00 output.
• Moonshot's own table beats Claude Opus 4.8 and GPT-5.5 on coding and agentic rows (Terminal-Bench 2.1, DeepSWE, BrowseComp) but concedes the overall crown to Fable 5 and GPT-5.6 Sol; Artificial Analysis placed it fourth at launch.
• This is datacenter-class open weight: 1.56TB targeting 64+ accelerator supernodes. Wait for community quants if your rack is a desktop.

Sources: Kimi K3 model card (Hugging Face), Kimi K3 License, Moonshot AI, Kimi K3, Tom's Hardware, Kimi K2 License, Glitchwire