kimi k3 is the largest open-weight model ever — and it tops the frontend code arena
Moonshot AI released Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model that leads the Frontend Code Arena at 1,679 points, ahead of Fable 5 and GPT-5.6 Sol. API pricing is $3/$15 per million tokens. Full weights drop July 27.
Beijing-based Moonshot AI launched Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter Mixture-of-Experts model that is, by a wide margin, the largest open-weight model ever released. full weights drop July 27 under a Modified MIT license. it's available now via API and inside the Kimi app.
the model architecture: 2.8T total parameters, 16 experts active per request out of 896, one-million-token context window, native image and video understanding. two variants shipped: K3 Max for chat and agent tasks, and K3 Swarm Max for large-scale parallel processing.
what it does on code
Kimi K3 took the number-one spot in the Frontend Code Arena — a live, blind developer evaluation — at 1,679 points. Fable 5 sits at 1,631. GPT-5.6 Sol at 1,618. K3 ranked first in six of the seven frontend domains tested, landing second only in Gaming, where Fable 5 still leads.
the Frontend Code Arena is a human-rated evaluation where developers submit real tasks and vote on which model's output they prefer — closer to how you actually use a coding tool than most academic benchmarks. a #1 rank there is meaningful in a way that synthetic evals often aren't. that said: K3's full weights aren't public until July 27, so independent community testing hasn't happened yet. the benchmark result comes from Moonshot's own deployment.
the price break
API pricing is $3 per million input tokens and $15 per million output tokens. that undercuts most Western frontier models — Claude Opus 5 is $5/$25, GPT-5.6 Sol is higher still. for teams running K3 via API the cost delta is material. for self-hosting, the 2.8T parameter size means you need serious hardware — realistically H100 or H200 clusters, not consumer GPUs.
the open-weight moment
the gap between open-weight and closed-source frontier models has functionally closed at the top of the coding leaderboard. DeepSeek started the pattern; Kimi K3 continues it at a larger scale. when K3 weights land on July 27, the community will be able to fine-tune, distill, and deploy the model without going through Moonshot's API. that's a different kind of leverage than API access alone.
the geopolitical layer: Kimi K3's release prompted reports that the Trump administration is revisiting a push to restrict Chinese AI models in the U.S., citing cybersecurity concerns. once weights are publicly downloadable, enforcement becomes nearly impossible — the same dynamic that made DeepSeek impossible to un-release. for developers evaluating K3 for production use, the policy uncertainty is real; for personal or research use, the weights will exist regardless of what happens in Washington.
what to watch on july 27
when the full weights drop, the first things worth looking at: (1) do independent coding benchmark results match Moonshot's claims on the Arena, (2) what does community fine-tuning surface about the model's actual coding reasoning versus its raw output quality, (3) can anyone get meaningful throughput on consumer-accessible hardware via quantization. the next week of community testing will tell you more than the launch numbers do.