Tencent's Hy4 Tops CWE-Bench-AA at 76%
Tencent's Hy4 ranks first on CWE-Bench-AA with 76%, six points ahead of DeepSeek V4.1 Flash 1017 (max).
Tencent's Hy4 ranks first on CWE-Bench-AA with 76%, six points ahead of DeepSeek V4.1 Flash 1017 (max).
A shared thread describes exporting, quantizing, and running every transformers model end-to-end on any backend with identical inputs and outputs as a drop-in r…
TypeSafe AI's Jev processes a photo in 94ms, roughly 30x faster than Sol, and delivers about 50x as many judgments for the same cost.
Anthropic's Engineering Blog shares 10 recommendations for how to design, run, and report safety evaluations that determine whether a model is ready to deploy.
TRL v1.18.0 makes AsyncGRPOTrainer stable, running generation and training simultaneously on separate GPUs instead of alternating, with documentation on usage a…
An AWS team published a walkthrough on building a self-service LLM gateway using Amazon Bedrock and LiteLLM.
LiteLLM introduced a Private Registry available to customers with an existing confidentiality agreement, offering early fixes for the most severe vulnerabilitie…
Google adds Google Wallet to Gemini's Personal Intelligence connected apps, enabling personalized finance answers with spending charts, subscription detection,…
Perplexity has introduced Teammates in its Mac app, available to Max subscribers. These agents can work individually or together to handle tasks such as QA, cus…
Requesty shared usage data showing that 68% of active teams used two or more models last week, 60% used two or more labs, 62% used two or more providers, and ab…