Kimi K3 just dropped. 2.8 trillion parameters, open-weight LLM model from Moonshot AI. First open-source model in the 3-trillion-parameter class. Built for long-horizon coding, knowledge work, and deep reasoning. On paper it looks absolutely unhinged.
Coding
K3 is a coding-first model. Don't expect it to write a novel. Its creative writing is serviceable, but it's not one of its strengths. In fact, K2.6 seems better at writing technical reports. If your workflow revolves around prose, there are better options. If you need to write code, however, K3 is exceptionally capable.
Frontend
Frontend development is where K3 truly stands out. It consistently delivers impressive results out of the box while showing a strong sense of creativity in UI/UX design. This has been true for K2 as well, but K3 takes it to another level. It can build beautiful, 3D animated websites with remarkable ease. Its repository navigation is also exceptionally solid, which makes a significant difference when working on larger projects.
As of today, I think it's the strongest frontend model available.
Backend and Security
K3 doesn't appear to focus as aggressively on security considerations and edge-case testing as GPT-5.6. When it comes to security-sensitive code or large-scale backend systems, I wouldn't be comfortable letting it operate without close supervision. Its actual greatness lies in frontend development and creative prototyping.
Token Burning
One repeating complaint from Kimi K3 users is that its token consumption is out of control. The $100 subscription tier gives you maybe four full sessions before you hit the weekly limit. K3 burns through tokens at a staggering rate for tasks that other models could complete using a fraction of the budget.
Kimi models have had a tendency toward lengthy reasoning loops ever since versions 2.6 and 2.7. While the deeper thinking is valuable for complex problems, it often overthinks relatively simple tasks, consuming far more tokens than necessary. With K2.5, you could build three complete full-stack SaaS applications without coming close to the usage limits. With K3, those savings are much harder to justify when the reasoning loops continue far longer than they need to.
Even so, it still delivers excellent value for the price and continues to punch well above its weight.
The Harness Matters
Though one thing to note is that the harness you wrap around the model matters alot. The default system prompts from Kimi's own tools may be inflating token usage. If you control that for your use case, the improvements can be drastic. Trying a different harness is recommended.
Verdict
Kimi K3 is genuinely smart. It's arguably the best model for frontend right now. It thinks ahead, generates beautiful code, and competes with models that cost way more. The token economics need work and it's not a general purpose model. But if you need a coding specialist you can run on your own infrastructure, this is it.
Its true value is in using it as a specialist tool. Use it when you need its reasoning and execution capabilities. For everything else, there are cheaper options that will get the job done.