What Is Kimi K3? Moonshot AI's 2.8-Trillion-Parameter Open Model

Moonshot's Biggest Open Release Yet

Kimi K3 is a frontier-scale AI model from Moonshot AI, the Beijing lab behind the Kimi chatbot and the earlier Kimi K2 model family. Announced July 16, 2026, Moonshot describes it as the world's biggest open-source model: 2.8 trillion total parameters, likely arranged as a mixture-of-experts architecture, with a roughly 1-million-token context window. The company followed the announcement with full open weights on July 26, 2026, published on GitHub and Hugging Face.

It's natively multimodal, handling text and images, and comes with Kimi Code, a terminal coding agent whose Agent Swarm feature can coordinate up to 300 sub-agents working in parallel on a single task. It's the direct successor to Kimi K2.7 Code, continuing Moonshot's rapid iteration through the K2 line earlier in 2026.

One Real Benchmark Win, Many Vendor Claims

That gap between a single narrow win and sweeping performance claims isn't unusual for a major model launch, but it's worth holding onto: a large parameter count signals scale, not automatically better real-world results. Test it against models you already trust on your own workload before switching.

China's Open-Weight Lead Keeps Widening

Kimi Code's Agent Swarm can coordinate hundreds of sub-agents on a single coding task.

Kimi K3 lands amid Moonshot's broader 2026 comeback story: after Kimi's consumer popularity dipped through 2025, the company pivoted hard to open-sourcing frontier-adjacent models, rebuilding its developer reputation and raising roughly $2 billion at a $20 billion valuation. Independent testing outside this launch has placed the K2 line as high as 4th overall on the Artificial Analysis Intelligence Index — ahead of Claude Opus 4.8 — making Moonshot arguably the leading open-weight AI lab in the world heading into K3's release.

What 2.8 Trillion Parameters Actually Means to Run

A parameter count that large is easy to quote and hard to picture in practice, so it's worth translating into what it means for anyone actually trying to run the model. Kimi K3 is almost certainly a mixture-of-experts (MoE) architecture, following the pattern of the earlier K2 line — meaning the full 2.8T parameters aren't all active on every token. Instead, a routing layer selects a smaller subset of "expert" sub-networks per request, so the effective compute cost per token is a fraction of what the total parameter count suggests. That's the only reason a model this size is usable at all outside a hyperscaler's own data center; a fully dense 2.8T model would be impractical for almost anyone to serve.

Even with MoE routing, storing the full weight set for self-hosting still requires terabytes of fast storage and multiple high-memory GPUs working in parallel, which is why most teams outside large labs access K3 through Moonshot's hosted API or a third-party router rather than downloading and serving the raw checkpoint themselves. The open-weight release matters most to researchers studying the architecture, companies with strict data-residency requirements willing to invest in the infrastructure, and cloud providers who package it into a managed endpoint for everyone else.

Agent Swarm's 300-sub-agent ceiling works on a similar principle of controlled parallelism: rather than one long-running agent working through a task sequentially, a coordinator process splits the work into smaller units and dispatches them to sub-agents that can run concurrently, then merges the results. This is genuinely useful for large, parallelizable coding tasks — refactoring across hundreds of files, or running a broad test suite — but it doesn't help with tasks that are inherently sequential, where each step depends on the previous one's output. Knowing which kind of task you're handing it is the difference between a swarm finishing in minutes and one that spends most of its time waiting on itself, since coordination overhead between sub-agents grows with every added layer of dependency.

Where Kimi K3 Fits

Frequently Asked Questions