Kimi K3: 2.8 Trillion Parameters, and They're Giving It Away
Moonshot shipped Kimi K3 today. 2.8 trillion parameters. The largest open-weight model anyone has ever released, from anywhere. Hacker News put it at the top of the front page within hours and it hasn't moved.
Here's what it actually is. A mixture-of-experts model with a 1M-token context window and native multimodal understanding, built on two pieces of architecture worth knowing about. Kimi Delta Attention is a hybrid linear attention scheme that gets up to 6.3x faster decoding once you're deep into million-token contexts. Attention Residuals buy roughly 25% more training efficiency for under 2% extra cost. That second number is the one to stare at. Moonshot didn't just scale up, they found a cheaper way to climb.
The benchmarks back it. Moonshot claims overall intelligence second only to Claude Fable 5 and GPT-5.6 Sol. K3 scored 1687 on GDPval-AA v2, which puts it ahead of Claude Opus 4.8 Max. On AA-Briefcase, a benchmark built specifically for long-horizon agentic knowledge work, it hit 1527 and came second only to Claude Fable 5 Max. An open model is now trading blows with the closed frontier on the exact workload that matters for agents: long tasks that don't finish in one turn.
The agent plumbing is there too. Automatic context caching, tool calls, JSON mode, structured output via JSON Schema, plus two new tricks: tool choice constraints and dynamically loaded tools. That last one matters more than it sounds. If your agent can load tools at runtime instead of declaring everything up front, your context budget stops being a hard ceiling on how many tools the agent can reach.
It's live right now on kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open weights land by July 27. That gap is the interesting part. Moonshot gets ten days of API revenue and press before handing the whole thing to anyone with enough GPUs. And the same day K3 shipped, Open Interpreter and LM Studio both launched agent harnesses built specifically for open models. Nobody coordinated that. The tooling showed up because the model was coming.
https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
← Back to all articles
Here's what it actually is. A mixture-of-experts model with a 1M-token context window and native multimodal understanding, built on two pieces of architecture worth knowing about. Kimi Delta Attention is a hybrid linear attention scheme that gets up to 6.3x faster decoding once you're deep into million-token contexts. Attention Residuals buy roughly 25% more training efficiency for under 2% extra cost. That second number is the one to stare at. Moonshot didn't just scale up, they found a cheaper way to climb.
The benchmarks back it. Moonshot claims overall intelligence second only to Claude Fable 5 and GPT-5.6 Sol. K3 scored 1687 on GDPval-AA v2, which puts it ahead of Claude Opus 4.8 Max. On AA-Briefcase, a benchmark built specifically for long-horizon agentic knowledge work, it hit 1527 and came second only to Claude Fable 5 Max. An open model is now trading blows with the closed frontier on the exact workload that matters for agents: long tasks that don't finish in one turn.
The agent plumbing is there too. Automatic context caching, tool calls, JSON mode, structured output via JSON Schema, plus two new tricks: tool choice constraints and dynamically loaded tools. That last one matters more than it sounds. If your agent can load tools at runtime instead of declaring everything up front, your context budget stops being a hard ceiling on how many tools the agent can reach.
It's live right now on kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open weights land by July 27. That gap is the interesting part. Moonshot gets ten days of API revenue and press before handing the whole thing to anyone with enough GPUs. And the same day K3 shipped, Open Interpreter and LM Studio both launched agent harnesses built specifically for open models. Nobody coordinated that. The tooling showed up because the model was coming.
https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
Comments