Chinese company Moonshot AI, backed by Alibaba, has released Kimi K3 — a large language model with 2.8 trillion parameters. That makes it the largest open model in the world, far bigger than DeepSeek V4 Pro (1.6 trillion) and other open alternatives. Kimi K3 has a 1-million-token context window, native visual understanding and a persistent reasoning mode called "thinking mode". The company claims the model delivers performance comparable to leading proprietary systems from Anthropic and OpenAI.
What happened
Kimi K3 landed shortly before the World Artificial Intelligence Conference (WAIC) 2024 in Shanghai. The model is built on two of Moonshot AI's in-house architectural innovations: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which the company describes as a replacement for residual connections that delivers steady scaling gains. Both techniques were previously published by the Moonshot team as open research on GitHub. Full model weights are slated for release on 27 July. Kimi K3 is already available for testing at kimi.com, with no credit card required to sign up.
Why it matters
The Kimi K3 release is a major event in the global AI race and a turning point for the open-source movement. It is also an impressive comeback for Moonshot AI, whose market position weakened after DeepSeek's rapid rise. The model posted strong numbers across benchmarks: on GDPval-AA v2 it took third place, behind only Claude Fable 5 Max and GPT-5.6 Sol Max. On the private AA-Briefcase benchmark K3 climbed to second, and on BrowseComp, which measures information retrieval, it hit a record 91.2 out of 100. In one demo, K3 designed a physical chip on its own in 48 hours using open EDA tools, pointing to serious autonomous agentic capability.
What it means in practice
- New development options: Kimi K3 is compatible with the OpenAI SDK, which simplifies integration for developers already using OpenAI or Anthropic tooling. That lowers the barrier to building products on top of a powerful open model.
- API savings: Kimi K3 API pricing is $3 per million input tokens and $15 per million output tokens, with cached input tokens at $0.30 per million. That is competitive against Western alternatives while offering performance on par with leading models.
- Potential for autonomous agents: The chip design demo shows K3 can sustain long, multi-step autonomous work. That opens new ground for complex agentic systems capable of engineering tasks with minimal human intervention.
- Flexibility for enterprises: An open model at this scale and performance level gives companies room to fine-tune, self-host and build proprietary systems without locking themselves into API contracts with big players like OpenAI or Anthropic.
- A growing coding ecosystem: Moonshot AI keeps investing in its coding ecosystem, including Kimi Code CLI, which has received significant updates. That widens the options for developers using AI agents to automate programming work.