LaunchAI ModelsJuly 28, 2026
Moonshot AI releases Kimi K3, 2.8 trillion parameter open-weight model

Kimi K3 features 2.8 trillion parameters, 1 million token context, native multimodal input, and Kimi Delta Attention for up to 6.3x faster decoding. It tops Code Arena and scores 71.6 on Perplexity's DRACO benchmark, surpassing GPT-5.6 Sol and Claude Fable 5. Moonshot open-sourced AgentENV and FlashKDA.
15 sources
unsloth/Kimi-K3-GGUFhuggingface.co
Code Arena now measures fullstack capabilities! View overall rankings across AI models on...x.com
Kimi K3 Architecture Notessebastianraschka.com
China Cuts AI Gap as Moonshot Shows Zhipu Isn’t One-Off, BI Saysbloomberg.com
[AINews] not much happened todaylatent.space
Chinese startup Moonshot AI unveils Kimi model it says rivals OpenAI, Anthropiccnbc.com
Kimi K3, and what we can still learn from the pelican benchmarksimonwillison.net
Fable, Sol Pro, Kimi K3: "write me a short but good poem using the Odyssey as a basis, think Tennyso...bsky.app
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Waymo says AI projects aren't ready until evals pass
- Reddit user reflects on rapid AI progress from GPT-5 era to local Qwen3.6-27B
- 5 startups address enterprise AI agent interoperability, trust, and audit gaps
- Katy George on the AI question every team should ask
- Claude experiences service outage