Kimi K3 launches as a 2.8T open-weights model with 1M context and $3 input pricing

The Batch (Andrew Ng) · rss · 2026-07-22

Main headline: Kimi K3

Moonshot released Kimi K3, a 2.8 trillion-parameter open-weights model with native vision, a 1 million-token context window, and a sparse MoE design that activates 16 of 896 experts.

The company says two architecture changes, Kimi Delta Attention and Attention Residuals, improve scaling efficiency by about 2.5x versus K2.
API pricing is live now at $3 / million input tokens and $15 / million output tokens; cached coding input can drop to $0.30 / million tokens.
Full weights are scheduled for release by July 27, 2026.
Moonshot demoed K3 on autonomous GPU compiler development, chip design in 48 hours, and automated video editing, while acknowledging it still trails Claude Fable 5 and GPT 5.6 Sol in head-to-head tests.

Other items in the same issue

Thinking Machines released Inkling, a 975B-parameter open-weights MoE model with 41B active parameters, trained on 45T tokens across text, images, audio, and video.
Inkling emphasizes controllable thinking effort and matches Nvidia’s Nemotron 3 Ultra on agentic coding tasks with about one-third of the tokens.
The EU ordered Google to open Android more to rival AI services and share anonymized search data, after finding non-Google agents could not function at Gemini’s level on Android.
Nvidia launched Nemotron 3 Embed, a family of open embedding models; the 8B model tops RTEB at 78.5%, and the 1B variants are optimized for production and Blackwell GPUs.
Google renamed NotebookLM to Gemini Notebook.
Hugging Face said it used GLM 5.2 to help analyze a weekend-long AI agent intrusion that its commercial frontier models refused to inspect.

Original post →

Kimi K3 launches as a 2.8T open-weights model with 1M context and $3 input pricing

More from Companies & People