Kimi K3 Weights Are Public. I Still Can't Run Them on My Laptop.

Ten days ago I wrote that Kimi K3 had changed my mind. Moonshot's model was live on the API, I was running it inside Hermes, and the open weights were still on a countdown for July 27.

That countdown is over. The weights are up on Hugging Face under moonshotai/Kimi-K3. The technical report is public on GitHub. And the infrastructure pieces Moonshot used to train the thing (MoonEP, FlashKDA, and AgentEnv) are open too.

I still pay Anthropic $200 a month for Claude. I still run K3 in my agent setup. Today those two facts sit next to a third one: anyone with enough hardware can download the same model I have been calling through an API.

What actually shipped

Per Moonshot's own model card and tech blog, Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model. It activates 16 of 896 experts per token. The Hugging Face card lists 104 billion activated parameters. Context window is 1,048,576 tokens. It has native vision through MoonViT-V2. The weights were trained with quantization-aware training, shipping as MXFP4 weights with MXFP8 activations.

Moonshot calls it the world's first open 3T-class model. That is their framing. What is checkable without the marketing line is simpler: 2.8T total parameters, open weights, public license file, public tech report.

The API launch was July 16. The company said full weights would land by July 27. Hugging Face shows the Kimi-K3 repo live, with downloads already in the thousands and community activity on the model card the same day.

Moonshot also published training-side tools alongside the weights. Per their Open Day post, that includes MoonEP (expert-parallel communication), FlashKDA (kernels for their attention design), and AgentEnv (sandboxing for agent training). Those are useful if you train models. For people like me who mostly call models, the headline is still the checkpoint itself.

Free to download is not free to run

This is the part a lot of "open weights dropped" headlines skip.

Moonshot's own tech blog says it recommends deploying K3 on supernode setups with 64 or more accelerators in a high-bandwidth domain. Secondary hardware writeups (explainx.ai, TechTimes) put the practical floor in multi-GPU / multi-node territory, not a desktop or a single workstation GPU. The native pack is MXFP4, which is smaller than full 16-bit, but "smaller than enormous" is still enormous.

I am not downloading this onto my laptop. I am not pretending I am. For a regular person building with agents, the realistic paths on day one are the same ones that already existed: Moonshot's API, or a host that already racked the GPUs.

On that second path, explainx.ai reported that Together AI and Modal both announced day-0 hosted access timed to the drop. Moonshot's GitHub README points self-hosters at vLLM, SGLang, and TokenSpeed recipes.

So "open" here means the weights are public and inspectable. It does not mean the average builder suddenly got a free frontier model on consumer hardware.

What the license actually says

I read the LICENSE file on the Hugging Face repo. It is the Kimi K3 License, not a blank check.

You can use, copy, modify, fine-tune, and ship products under it, with the usual copyright notice. Two commercial gates matter if you are building a business on top of it:

If you run a "Model as a Service" business (giving third parties real control over inference or fine-tuning, not just an end-user feature that happens to use the model), and you and your affiliates clear more than $20 million in aggregate revenue over any consecutive 12 months, you need a separate agreement with Moonshot before commercial MaaS use.

If a commercial product or service using the software (or derivatives) has more than 100 million monthly active users, or more than $20 million in monthly revenue, "Kimi K3" has to be displayed prominently in the product UI.

Internal use is carved out. So is use through Moonshot's own products or certified inference partners. Everything is AS IS, no warranty.

That is more open than a closed API. It is not "do whatever you want forever with zero strings." Anyone planning a commercial deployment should read the file themselves rather than trusting a social media summary that just says "MIT."

Price on the API, next to Claude

I am still mostly an API user, so the price table still matters more to my wallet than the download button.

Per Moonshot's platform and the K3 tech blog, Kimi K3 API pricing is $3.00 per million input tokens, $15.00 per million output tokens, and $0.30 per million on cache-hit input. Pricing is flat across the full 1M context. Moonshot says its official API sees cache hit rates above 90% on coding workloads.

Per Anthropic's Fable pages and pricing docs, Claude Fable 5 is $10 per million input and $50 per million output, with cache hits at $1 per million input tokens on the standard table.

I am not going to turn that into a winner announcement. Moonshot's own blog says K3 still trails Claude Fable 5 and GPT-5.6 Sol on overall performance. Different models, different harnesses, different days. What I can say without guessing is the list prices: K3's published API input and output rates are lower than Fable 5's published rates, and both companies publish those numbers in public docs.

What this changes for someone who already uses it

Before today, if I wanted K3 I went through Moonshot's API, the Kimi apps, or whatever host had already wired it up. After today, the trained parameters are also a public artifact. Researchers can inspect them. Companies with real GPU budgets can self-host. Third-party hosts can compete on serving price without waiting for Moonshot's permission to touch the weights.

For me personally, the day-to-day loop does not magically change overnight. I still pay $200 a month for Claude. I still run K3 in Hermes when the job fits. The new fact is narrower and still big: the model I have been using through a meter is now a downloadable checkpoint under a published license, with a tech report next to it.

That is what dropped on July 27. Not a press release about a future open release. The weights.