Moonshot AI launched Kimi K3 on 16 July 2026: 2.8 trillion parameters, a 1M-token context window, and the top spot in a frontend-code benchmark. Open weights are due July 27.
On 16 July 2026, Chinese company Moonshot AI released Kimi K3 — the largest open-weight model to date, with roughly 2.8 trillion parameters. It succeeds the Kimi K2 family, with full weights due July 27 under a Modified-MIT-style license. It's a notable milestone in the 2026 AI race.
The headline numbers
- 2.8 trillion parameters in a Mixture-of-Experts (MoE) design.
- A 1-million-token context window — enough to load an entire codebase or hundreds of pages into a single prompt.
- Native vision and always-on reasoning.
- Aimed at long-horizon coding and agent workloads.
What is Mixture-of-Experts, and why it matters
2.8 trillion sounds enormous, but MoE keeps the real cost far lower. Instead of activating the whole network for every word, the model fires only 16 of its 896 "experts" per token — about 1.8%. Think of a large hospital: you don't see a hundred doctors; the system routes you to the few specialists that fit your case. Moonshot uses a "Stable LatentMoE" framework that improves scaling efficiency by roughly 2.5× over Kimi K2 — converting compute into intelligence more effectively.
Benchmark performance
In blind developer testing, Kimi K3 ranked #1 in the Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5. It jumped 17 places from K2.6 (#18 → #1) with a 76% pairwise win rate — versus 63% for Claude Fable 5 and 58% for GPT-5.6 Sol. Architecturally, K3 introduces KDA — a hybrid linear-attention mechanism interleaving linear- and full-attention layers in a 3:1 ratio — plus "Attention Residuals" for more efficient scaling.
How to try it
Kimi K3 is already available through Kimi products and an OpenAI-compatible API — meaning a project using the OpenAI SDK can point to it with almost no code change. A minimum $1 top-up unlocks it; your cumulative top-up sets your rate limits. Listed pricing: $0.30 per million cache-hit input tokens, $3 per million on cache misses, and $15 per million output tokens. Once weights publish (expected July 27), enterprises can self-deploy rather than depend on an API. Note: at the time of writing, a full checkpoint and model card were not yet public — wait for the official release before planning to self-host.
The big picture: China's open-model wave
Kimi K3 doesn't stand alone. It's part of a strong wave of Chinese open-weight models (GLM, DeepSeek, Qwen…) narrowing the gap with top U.S. frontier systems. The 2026 trend is clear: the gap between open and premium commercial models on everyday work is shrinking to single-digit percentages, while costs run several times lower. For engineering teams in Vietnam, it's one more powerful, cost-efficient option to weigh when building AI products — as long as you read the license terms carefully before shipping commercially.
References
Comments (1)
fsdfsdf