Kimi K3's weights are finally public, and they weigh 1.4 terabytes
Moonshot AI has published the full weights and technical report for its 2.8-trillion-parameter Kimi K3, along with parts of the infrastructure needed to run it. The claim worth checking: 2.5 times more intelligence per unit of compute.
Moonshot AI has completed the release it started talking about in mid-July. The Chinese lab published the weights for Kimi K3 on Hugging Face on Monday, together with a technical report on GitHub and, more unusually, several pieces of the plumbing needed to actually run the thing: high-performance attention kernels, a library for the communication between experts inside the model, and tooling for operating AI agents at scale. Within hours it was the most-discussed item on Hacker News.
A quick translation of the specs. Kimi K3 is a mixture-of-experts model, meaning it contains many specialised sub-networks and routes each piece of text to only a handful of them. On paper it has 2.8 trillion parameters, the tunable numbers that hold what a model has learned, but it activates just 16 of its 896 experts per token, so running it costs far less than the headline number suggests. It handles a context window of one million tokens, roughly the length of a very long book held in mind at once, and it understands images and video natively rather than through a bolt-on. Moonshot’s central claim in the report is architectural: a new attention design it calls Kimi Delta Attention plus a restructured expert framework deliver about 2.5 times more intelligence per unit of compute than its previous generation. On Artificial Analysis’s composite Intelligence Index the model scores 57, close to GPT-5.6 Sol, and it currently tops the Frontend Code Arena leaderboard.
Then there is the practical catch, which is the part most coverage skipped: the download is around 1.4 terabytes. Open weights mean you are legally and technically free to run the model yourself, but freedom is not the same as feasibility. Serving K3 at usable speed needs a rack of datacentre GPUs, not a laptop, so for most people the realistic route is a hosted provider rather than a local install. The honest caveats also include capability gaps. Independent testing by the UK’s AI Security Institute found K3 trailing frontier US models by a wide margin on cyber-exploit tasks, and it lags on hard mathematics too. Both patterns are consistent with distillation, where a smaller model is trained on the outputs of a stronger one, an accusation Chinese labs face routinely and that Moonshot has not addressed directly.
What this means for you: If you are just getting oriented, the useful lesson is what “open weights” actually buys. It does not mean free, and it rarely means local. It means nobody can switch the model off, raise its price, or change its behaviour underneath you, which matters a lot if you are building something on top of it and not much if you are chatting in a browser. If you already work with models, K3 is worth a look for front-end code generation specifically, where it currently leads, and worth avoiding for security work and heavy maths until independent scores improve. Everyone else can watch the price effect instead: each credible open release from China puts downward pressure on what Western labs can charge, and that reaches your bill eventually.
Sources
Microsoft's new security model is small on purpose, and that is the interesting part
MAI-Cyber-1-Flash handles about 90 percent of vulnerability-hunting tasks and hands the rest to GPT-5.4. Microsoft says the combination scores 96 percent on CyberGym at half the cost of its previous setup.
