OpenAI’s safety system is already cutting off API responses mid-task
The New Stack 1 month ago 35
OpenAI’s safety system has been stopping API outputs mid-task for some early users when models are restricted under its Preparedness Framework.
Model Evaluation
204 summarised stories about Model Evaluation, each linking back to the original source. Browse all topics →
Key players
Friday, 11 September 2026
Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages
MarkTechPost 1 month ago 52 ● 3 sources
Cohere released North Small Translate, an open-weight sparse MoE machine translation model for 50 languages. It reports an average WMT26 all-languages score of 83.6, with 25B active parameters out of 218B total. The model becomes available to use via Cohere’s API (free until rate limits), or to self-host non-commercially using released checkpoints, with commercial licensing options.
DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation
Apple Machine Learning Research 1 month ago 44
DiscoSign introduces a discourse-aware text-to–sign-language gloss translation framework that targets discourse phenomena ignored by sentence-level systems. It improves spatial consistency and entity tracking on experiments using sentence-level and discourse-level datasets. The method adds modules for spatial coreference, question-answer clause handling, and concept-gloss consistency plus discourse-level evaluation metrics that replace sentence-only quality measures.