· 5 min · Tools
Kimi K3 Tops the Coding Charts — and Open-Sources the Weights Next Week
Moonshot's 2.8-trillion-parameter model is the largest ever released with open weights, and it's beating Fable 5 on an independent coding leaderboard — but it dropped the cheap-Chinese-model playbook, and the weights it's named for aren't downloadable yet.
On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that is the largest ever put out with open weights — the downloadable files that let anyone run a model on their own hardware instead of renting it through an API. Within a day it sat at the top of an independent coding leaderboard, ahead of Claude Fable 5 and OpenAI's GPT-5.6 Sol. Two things about that sentence are shakier than they sound: the leaderboard covers only front-end code, and those "open weights" can't be downloaded yet.
Strip away the superlatives and Kimi K3 is the clearest sign yet that the gap between China's open models and America's closed frontier has shrunk to a rounding error on some tasks. But it's also a break from the script those Chinese models have followed for two years. K3 isn't cheap, and — as of this writing — it isn't actually open. Both of those deserve more of a developer's attention than the leaderboard rank does.
What it is
K3 is a mixture-of-experts model: of its 2.8 trillion parameters, spread across 896 "experts," only 16 fire on any given token (the unit of text a model reads and writes). That's the trick that makes a model this enormous affordable to serve — most of it sits idle for any given word. It uses a new attention design Moonshot calls Kimi Delta Attention, carries a 1-million-token context window (how much text it can consider at once), and understands images natively. For now it's reachable through Moonshot's website and API.
The leaderboard, and how much to trust it
K3 debuted at number one on the Frontend Code Arena with a score of 1,679, a 17-place jump from its predecessor and ahead of both Fable 5 and Sol, winning six of the arena's seven front-end categories (it trails Fable 5 only in gaming). This is a real independent signal, not a vendor slide: the arena ranks models by human preference votes on head-to-head outputs — a stronger footing than some earlier Chinese coding-model claims, where only one benchmark score had been checked by anyone but the maker.
It holds up beyond that one venue, too. The third-party evaluator Artificial Analysis scores K3 among the top three across six coding benchmarks, leading on two and trailing Sol on a terminal-tasks test by half a point. So it isn't a single-leaderboard fluke. But "best at front-end preference tests" is not "best coding model," and it shouldn't be read as one — preference voting rewards output that looks right, front-end work is where that bias is strongest, and public leaderboards can be optimized for. The developer Simon Willison, testing it hands-on, called K3 a clear step up from prior Kimi versions but flagged a practical catch: it ships with a single reasoning setting, "max," so it thinks hard — and expensively — about everything. Which brings up the price.
The twist: it isn't the cheap option
For two years the story of Chinese open models was frontier-adjacent quality at a fraction of the price. K3 walks away from it.
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Kimi K2.6 (its predecessor) | $0.95 | $4 |
| Kimi K3 | $3 | $15 |
| Claude Sonnet (for reference) | $3 | $15 |
Moonshot priced its flagship at exact parity with a mid-tier US frontier model, roughly triple what the last Kimi cost. And because the only reasoning mode is the token-hungry "max," the real cost per task runs above the sticker rate — one of Willison's simple tests burned about 16,600 output tokens, some 25 cents, on a single prompt. If the reason to look at a Chinese model was that it undercut Sonnet and Sol, K3 is not that model.
The other twist: "open" is still a promise
Moonshot says it will publish the full weights on Hugging Face — the standard hub for sharing models — by July 27, under a "Modified MIT" license. Permissive, but the "modified" is doing work; anyone planning to build on it should read the actual terms rather than assume plain MIT. And as of this writing, the repository is still empty.
That gap matters because open weights are the entire point of a story like this. Self-hosting for privacy, fine-tuning on your own data, running it air-gapped, and — not least — independently verifying those benchmark claims by running the thing yourself: none of that is possible until the files land. Until July 27 delivers, the largest open-weight model ever released is, in practice, just another paid API.
What it means for you
If you build front-end- or UI-heavy features, K3 is worth a genuine bake-off — its strength there is real and independently observed, not marketing. Don't switch for the price, though: at $3/$15 plus a mandatory "max" reasoning mode, it isn't undercutting Sol or Fable 5. And if self-hosting or fine-tuning is the whole reason you care, wait for July 27, confirm the weights are actually up, and read the license before you plan around them — openness you can't download is a roadmap item, not a feature. The lasting signal is directional: an open-weight model trading the top spot with the closed frontier, and pricing itself like the closed frontier, is the same convergence showing up when Meta started charging for Muse Spark. The tidy split between cheap-and-open and expensive-and-closed is dissolving.