Beijing-based Moonshot AI has published the full weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that it describes as the world's first open 3-trillion-class system and the largest open-weight AI model released to date. The weights dropped on July 26, 2026 — a day ahead of the company's stated July 27 target — landing 96 shards, a license and deployment instructions on Hugging Face, with cloud platforms including Together AI and Modal offering day-zero hosted access.
What's Inside K3
Kimi K3 is a mixture-of-experts (MoE) model, a design that keeps total capacity enormous while activating only a fraction of parameters per token. Of its 2.8 trillion total parameters, roughly 104 billion are active for any given token, with 16 of 896 routed experts firing alongside two shared experts. The context window is a full 1,048,576 tokens — just over a million — and the model accepts both text and images.
For scale, 2.8 trillion total parameters is about 75% larger than the previous largest widely used open model. That heft is a statement of intent: while much of the open-source community has chased efficiency and smaller footprints, Moonshot has gone in the opposite direction, releasing a genuinely frontier-scale system for anyone to download.
Strong, but Not the Crown
Moonshot positioned K3 aggressively, comparing it against the strongest proprietary systems and reporting benchmark wins over some of them on coding and agent tasks. Independent testing tells a more measured story. On the third-party Artificial Analysis Intelligence Index v4.1, Kimi K3 scored 57.1, landing as roughly the #3 model family and #4 tested configuration — close behind the leading proprietary systems that scored in the high 58s and high 59s, but not ahead of them.
The company's own launch materials are candid about this: K3 still trails the very strongest closed models overall, and several of its headline comparisons mix different agent harnesses, which makes apples-to-apples reading difficult. The honest summary is that Kimi K3 has narrowed the gap between open and closed frontier models to a slim margin — an achievement in itself — without unseating the leaders.
Open, but Not Easy
There is an important asterisk on the word "open." Downloading the weights is free, but running K3 is not trivial. Moonshot recommends supernode configurations with at least 64 accelerators for deployment — hardware far beyond a hobbyist rig or a single server. Open weights democratize access to the model's parameters; they do not democratize the compute required to serve it. In practice, most organizations will reach K3 through hosted providers rather than self-hosting.
Pricing on those hosted endpoints reflects a frontier product: $3.00 per million cache-miss input tokens, $0.30 per million cache-hit input tokens, and $15.00 per million output tokens, charged at one flat rate across the full million-token context.
Why It Matters
Kimi K3 sharpens two debates at once. The first is open versus closed. A downloadable model within striking distance of the best proprietary systems strengthens the case that open weights are a durable public good — giving researchers, startups and regulators the ability to inspect, fine-tune and audit frontier-class systems rather than treating them as sealed boxes.
The second is geopolitical. The release arrived amid intense friction over how China's labs are advancing despite US chip-export controls:
- A senior US science-policy official publicly accused Moonshot of running an internal distillation platform against a leading Western model using restricted Nvidia chips routed through Thailand.
- In response to the broader dispute, Anthropic's chief executive stated the company has never advocated a ban on open-weight models, calling non-dangerous open models a public good — while still pressing for chip-export enforcement, anti-distillation measures and mandatory safety testing.
Whatever the merits of those claims, the underlying signal is unmistakable: Chinese labs are shipping open frontier models at a cadence and scale that keeps them near the front of the pack, and doing so under significant hardware constraints.
The Takeaway
Kimi K3 is best understood not as a single benchmark victory but as a milestone in the openness of frontier AI. It proves that a 3-trillion-class system can be released to the public, hosted on day one across multiple clouds, and land within a few points of the best closed models. The catch — that serving it demands industrial-grade hardware — is a reminder that in 2026 the real bottleneck is no longer model access but compute. For a field increasingly defined by who controls the chips, that is the most consequential lesson of all.
