Moonshot has launched Kimi K3, a multimodal model with 2.8 trillion parameters and a 1-million-token context window, while the Schema harness enables frontier models to achieve 99% on the ARC-AGI-3 benchmark.
- Kimi K3 features optimizations for faster long-context decoding and agentic coding, with open weights scheduled for July 27.
- Google delayed Gemini 3.5 Pro to improve performance, particularly for coding tasks.
- Schema allows agents to write executable programs to test game mechanisms against reality in ARC-AGI-3.
- NVIDIA released three open embedding models for RAG and agentic retrieval, led by an 8B model ranking first on RTEB.
These developments highlight advances in long-context processing, benchmark performance, and specialized tooling for AI agents.