🛠️ Outils & Produits
vLLM v0.25.0 is now available, featuring 558 commits from 232 contributors. The release makes Model Runner V2 the default execution path for all dense models, building on quantized-model support from the previous version. (https://github.com/vllm-project/vllm/releases/tag/v0.25.0)
llama.cpp released commit b9969, which includes a Vulkan fix for the llama-cli breaking on longer prompt sizes for q4_0 quantized networks due to insufficient shared memory. (https://github.com/ggml-org/llama.cpp/releases/tag/b9969)
browser-use version 0.13.4 was released, adding an MCP server for CLI 3.0, fixing navigation readiness detection, and improving markdown extraction. (https://github.com/browser-use/browser-use/releases/tag/0.13.4)
crush version 0.84.1 was released, bumping the upstream Fantasy dependency to properly send session ID headers for session affinity. (https://github.com/charmbracelet/crush/releases/tag/v0.84.1)
🔬 Recherche
Long-Horizon-Terminal-Bench is a new benchmark for testing AI agents on long-horizon terminal tasks with dense reward-based grading, moving beyond simple, short-duration evaluations. (https://huggingface.co/papers/2607.08964)
Jet-Long is a research paper on efficient long-context extension for LLMs using Dynamic Bifocal RoPE, aimed at applications like retrieval-augmented generation and agentic workflows. (https://huggingface.co/papers/2607.07740)
OpenCoF explores a novel reasoning path for large models through video generation, using temporally connected frames as a distinct approach from traditional Chain-of-Thought. (https://huggingface.co/papers/2607.08763)
⚖️ Régulation & Société
Over 200 protesters marched to the offices of OpenAI and Google DeepMind, demanding stronger AI regulation and a pause in the AI race. (https://news.google.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?oc=5)
Experts are advocating for balanced AI regulation to unlock the potential of the digital economy. (https://news.google.com/rss/articles/CBMixgFBVV95cUxNYTJOd21JRDVNbUhEdHJwdzBjYVQ0bDMxaTNyTjNxcEcyTkFtaXhXaUJDYUJRWktDLEtnSURJZDROalVUYVpzdUNqbkVHSnFQN29FTnRpSDhtQ2E2X3JEQnkwRUQwNUNfd1poUHpwckI0SEdMSVNJZTdTWkZTLWVCWkdEMXl0cWlqX0RhYlNNcXpRenZmTjNZTXpial9YS09ybUdsOUlOMFcwTUQ1cFlkVlp4QnZFYlppdDU1T2xVTEdERVU5dWc?oc=5)
A law school is being highlighted as a model for good AI policy development. (https://news.google.com/rss/articles/CBMirAFBVV95cUxNS2RZR1llRGpnWVh5WXE0U2h0cXIxVTUtbDF6SGt3cWZ5TWVDbGtTc1dxdWV2NlFUbFB5X0V4cEhGOWRBMzUzTGwwVHBJRE9tUlFzeFlNeFlkdzhrMGM2LTRaZXVMdkxSWjdLcmRaWEsyX29IeVFXMWhyZldjSVFmT3U5ZW1TRklBQmsyNm9SQ2VVV1hPSUMwUE9Ddm5UaVBHdlJBbTlndzIwYTRW?oc=5)
🏢 Industrie & Business
OpenAI is hiring a dedicated product manager to build experiences for families, caregivers, and older adults, signaling a deeper push into household integration for ChatGPT. (https://techcrunch.com/2026/07/11/openai-bets-on-families-as-chatgpt-goes-deeper-into-households/)
💡 À retenir
vLLM's major v0.25.0 release makes Model Runner V2 the default for dense models, a significant backend change for anyone using the library for production inference. Meanwhile, public pressure for AI regulation is mounting, with protests at major lab offices.