Section
A single WSE-3T wafer holds 44 GB of SRAM; a trillion-parameter model needs terabytes. Cerebras shows a 59-wafer pipeline on open Kimi K3 and claims 1,000–2,000+ tokens a second.
Read
Frontier models and Codex agent workflows will land in Model Garden. Buys go through Marketplace and can draw down GCP commitments. Microsoft stays the primary cloud partner.
Read
Post-training on top of Kimi K3, a million-token context, a Highspeed variant 2–3× faster. On Terminal-Bench 4.0 it scores 65.7%. Full weights by 4 November.
Read
A separate Opus 5.5 app for Pro and Max plans in the US. A morning brief, background jobs, approval before mail and payment. Still a research preview on a waitlist.
Read
A small update to the open Flash model lifts coding and agent scores while cutting time and answer length. MIT weights, a million-token context.
Read
A US open model with 42 billion active parameters and a million-token context. Meta calls it the strongest US open-weight so far and shows it running on a pair of M5 Ultras.
Read
The new Gemini generation is tuned for science, code and long refactors. Inside Google it already helped Chrome and quantum chemistry. A robotics branch ships separately.
Read
Harvard finds agents lift code volume by 30%, but feature output does not rise: review swells by almost half.
Read
The new Muse Wonder and Muse Wonder Full line is aimed at personal and work agents: memory across sessions, long task focus, strong code. Access via Meta Model API and Muse Code.
Read
Researchers proposed a scheme where many AI agents jointly improve each other and their own strategies — a step toward systems that evolve with little human input.
Read