Cohere shipped North Small Translate, a 218B MoE (25B active) translation model scoring 83.6 on WMT26 All Languages and beating DeepL and Google Translate in its evals. Weights are on Hugging Face under CC BY-NC 4.0 for research use; enterprise access runs through RWS Language Weaver.
Google’s Developers Blog (Sept 11) describes autofinetune — agents that edit Tunix/Gemma training scripts, run SFT and GRPO on Cloud TPUs, and keep winning commits. Case studies cover FunctionGemma SFT and Gemma 3 1B math RL…
Mistral and Cloudera are integrating Mistral models into Cloudera’s hybrid platform for on prem, cloud, and air gapped inference, plus custom training on proprietary data so enterprises keep ownership of the resulting models…
OpenAI’s GPT Live 1 is live in the API: a full duplex voice model that listens and speaks at once, handles interruptions in one pass, and can hand deeper reasoning or tools to a backend like GPT 6 Astra. Front end voice is $0.05/min; telephony and broader voice options are included…
Claude Code 2.1.268 (Sept 10) restores third party Anthropic compatible endpoints that returned HTTP 400 since 2.1.265, caps hung WebFetch calls at 300s, and stops MCP/plugin errors from exposing tokens. Gateways can also push pricing: rates into managed settings so /cost matches the spend meter…
Replit’s Databricks integration is generally available, with native Lakebase support so Replit Agent can auto provision a managed DB on deploy. Apps can read live warehouse data and store app state under Unity Catalog permissions, lineage, and approval gates…
Spotted something we missed? Start a thread.