Moonshot Gave Away a 2.8-Trillion-Parameter Model. The Catch Is 1.4 Terabytes

On 27 July, the Beijing lab Moonshot AI released the full weights of Kimi K3, the largest open model anyone has handed out. ‘Open weights’ means you can download the actual model, run it on your own hardware, fine-tune it, and nobody sends you a bill per query. Compare that with OpenAI or Anthropic, where you rent access through an API and pay by the token forever.

K3 first appeared on 16 July as a hosted service. The weights followed on 27 July, all 96 shards of them, on Hugging Face under Moonshot’s own licence.

The numbers, and why they are silly

K3 carries 2.8 trillion parameters (the internal dials a model tunes during training). It uses a mixture-of-experts design, which means it only fires up a fraction of itself for any given token: 104 billion parameters active at a time, picked from 896 specialist ‘experts’. Context runs past a million tokens, and it takes text, images and video natively.

On the independent leaderboards it holds up. Kimi K3 topped WebDev Arena at 1,678 Elo, the first open model ever to lead that front-end coding board. On Artificial Analysis’s Intelligence Index it scored 57.1, roughly the third strongest family going, behind OpenAI’s GPT-5.6 Sol and Anthropic’s Fable 5, but ahead of plenty of things people pay handsomely for.

Now the catch

‘Free’ is doing some heavy lifting. The full download runs to about 1.4 terabytes. Running a 2.8-trillion-parameter model at any decent speed needs a rack of high-end GPUs that most people and most small companies simply do not have. So for the little guy, ‘open weights’ mostly means you can rent it cheaply from whichever host spins it up, or hit Moonshot’s own API at around $0.30 per million tokens for cache-hit input, rather than literally running it in your box room.

Even so, the direction is the story. A Chinese lab is now trading blows with the top US labs on quality, then giving the model away and letting anyone host it. That drags the floor price of frontier-grade intelligence towards zero and piles pressure on every company whose whole plan is charging you rent on a model you can never see. If you build software, it means genuine choice. You are no longer locked to two American vendors and their pricing whims.

Did you know: at roughly 1.4 terabytes, a single copy of Kimi K3’s weights would not fit on most laptops sold today, and would take the better part of a day to pull down on an average home connection.

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →

Sources

Get the free weekly stack: the AI tools and moves that matter, hype filtered out.Subscribe free →
Scroll to Top