Moonshot AI Launches Kimi K3 With Open-Weight Plans and 1M-Token Context

Moonshot AI announced Kimi K3, a 2.8 trillion-parameter native-vision model that the company says is available through its Kimi API and planned for a full weights release by July 27, 2026.

A large model with long-context positioning

In its Kimi K3 announcement, Moonshot AI describes the model as a 2.8T-parameter system with native vision capabilities and a 1 million-token context window. The company says Kimi K3 is available through the Kimi API, while also stating that full weights will be released by July 27, 2026.

The official Kimi API platform lists Kimi K3 as Moonshot’s latest flagship model and repeats the 1M-token context specification. The platform also publishes token pricing: $0.30 per million tokens for cache-hit input, $3.00 per million tokens for input, and $15.00 per million tokens for output.

Moonshot’s announcement says KDA prefill caching is part of how the company serves the model at competitive token prices despite its scale. That detail matters because long-context models can be expensive to run when large amounts of prompt content must be processed repeatedly.

Demand has tested serving capacity

The launch has already created operational pressure. The Associated Press reported that Moonshot temporarily paused new Kimi subscriptions after demand for K3 overwhelmed capacity. AP also cited an Omdia analyst who said the model is compute-demanding and challenging to serve at scale.

That report underscores a familiar issue for large AI systems: publishing or offering access to a frontier-scale model is only one part of deployment. Serving the model reliably, especially with long contexts and high user demand, requires substantial compute capacity and software optimization.

Open-weight plans could shift optimization work outward

Moonshot’s stated plan to release full weights is significant because open-weight availability can allow outside infrastructure providers, researchers, and developers to work on serving efficiency. Common approaches in the broader AI infrastructure market include quantization, batching, memory optimization, routing, and serving-system improvements, though Moonshot’s sources do not quantify how much cheaper Kimi K3 inference may become.

A July 19, 2026 episode page for Peter Diamandis’s “Moonshots” podcast, indexed by Podscan, confirms that Emad Mostaque joined a discussion about the Kimi K3 release, open-weight models, frontier labs, quantization, and AI infrastructure. However, the public source information available here does not independently substantiate specific numerical predictions about future cost reductions, so those claims should be treated as commentary rather than established fact.

Why it matters

Kimi K3 arrives at a moment when open-weight and API-served models are competing not only on benchmark performance, but also on context length, multimodal capability, availability, and serving cost. Moonshot AI is emphasizing scale, a 1M-token window, native vision, and future weight availability. At the same time, AP’s reporting shows that real-world demand can strain capacity soon after launch.

For developers and enterprises, the practical question is likely to be less about the headline parameter count alone and more about whether Kimi K3 can be served predictably at the published prices and latency levels required by production applications. Moonshot’s API pricing and caching approach provide an initial commercial framework, while the planned weights release could broaden the group of organizations able to test, optimize, and deploy the model outside Moonshot’s own hosted service.

The immediate takeaway is that Kimi K3 is both a major model release and an infrastructure test. Moonshot AI has put forward an ambitious long-context, native-vision system with published pricing and an open-weight timeline. The next phase will show how well the model can be served at scale, and how much optimization becomes possible once broader access to the weights is available.