🚀 Modèles & Releases
Moonshot AI has released Kimi K3, a 2.8T parameter open-weight multimodal reasoning model suited for complex coding and long-horizon agentic workflows. (https://openrouter.ai/models/moonshotai/kimi-k3) Meta has released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks that accepts text, images, video, audio, and PDFs. (https://openrouter.ai/models/meta/muse-spark-1.1) Kwaipilot has released KAT-Coder-Air V2.5, a flagship-level Agentic Coding model designed to autonomously handle entire issues or business workflows. (https://openrouter.ai/models/kwaipilot/kat-coder-air-v2.5)
🔬 Recherche
A new arXiv paper, "Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models," explores the interpretation of in-context learning as conditional inference. (https://arxiv.org/abs/2607.15277v1) Another paper, "RoboTTT: Context Scaling for Robot Policies," introduces a model and training recipe that scales visuomotor context to 8K timesteps. (https://arxiv.org/abs/2607.15275v1) A third paper, "MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators," applies reinforcement learning to align flow models with objectives. (https://arxiv.org/abs/2607.15273v1)
🏢 Industrie & Business
OpenAI has built an LLM super-hacker called GPT-Red, which it uses as a sparring partner to help its other models boost their defenses against cyberattacks. (https://www.technologyreview.com/2026/07/15/1140514/meet-gpt-red-an-llm-super-hacker-openai-built-to-make-its-models-safer/)
🛠️ Outils & Produits
Google has added personalized AI avatars to Google Vids, letting users create videos starring a digital version of themselves, alongside Gemini Omni-powered tools for generating and editing videos. (https://techcrunch.com/2026/07/16/google-vids-now-lets-you-star-in-your-own-ai-videos/) Roblox has launched a new "Build" feature in its mobile app that lets users generate basic games using a single text prompt. (https://techcrunch.com/2026/07/16/roblox-launches-an-ai-powered-game-creation-feature-in-its-mobile-app/) Google is expanding its AI Mode to allow linking and interacting with select apps, moving beyond just answering questions. (https://techcrunch.com/2026/07/16/googles-ai-mode-now-lets-you-link-and-interact-with-select-apps/)
⚙️ API & Dépréciations
Anthropic has made mid-conversation system messages available on Claude Fable 5, Claude Mythos 5, and Claude Opus 4.8 via the API, Amazon Bedrock, and Google Cloud, with no beta header required. (https://docs.anthropic.com/en/release-notes/api#july-15-2026) Weaviate released v1.38.5, which includes LSM store performance improvements and a fix for potential deadlocks in batch vectorization. (https://github.com/weaviate/weaviate/releases/tag/v1.38.5) Weaviate also released v1.36.22, which includes fixes for LSM segment cleanup, WAL recovery crash-safety, and search-time HNSW entrypoint repair. (https://github.com/weaviate/weaviate/releases/tag/v1.36.22) NVIDIA's TensorRT-LLM v1.3.0rc21 includes a deprecation notice for the AutoDeploy backend. (https://github.com/NVIDIA/TensorRT-LLM/releases/tag/v1.3.0rc21)
⚖️ Régulation & Société
A $325 million Senate bill has been proposed that includes provisions for AI regulation and a duplex allowance. (https://news.google.com/rss/articles/CBMimwFBVV95cUxPbWlyLS13LTlzQWJORU91ZlJIcHozVHVjejNkUzJRdnVfOENzSTdPdGFRRGJYOXlnd0NHUDNnam5ka2FweUQ2ci0wZmU0R0xVanR3eFBWVnl0bU5ZUFZqSUhYU1cyeDJ2dk95blF2OXNNNm84TktDNzF6VXQzc2lDZHROMWhBQTJjbU9JQWVRNmEwNmllNjlNcmlhdw?oc=5) The city of Cortez has adopted an AI policy. (https://news.google.com/rss/articles/CBMieEFVX3lxTE56ZkV6bkRGLVR6RldORFdwQkcyMkJqbGRJNTJxX193QkFlS0ZoQnBWak9mdzZyRE0wRm5zWXBqbm1kMWtiZ1VqbkVWSXVHMWR5R2kycVRfcjFMTnpGTWVLOXpyTDJFRE93aGFqVUdTVG8xOFA4YW1SbQ?oc=5) The CEO of Anthropic has donated $1 million to an AI regulation super PAC. (https://news.google.com/rss/articles/CBMiygFBVV95cUxOR1g4Ym9JR2NnLVZDMmd3eFlxUmdKd3JaSzIzZUV4c3NhSFV2STFVb0NZbnFPQ05ZNW1MVjNXZjVDcTUtQ2VVeF8tbHlIbFZMbFZYbjNVbXFkbF8xWVRxbjJXYXZuUWY2TFVEYzg0MG1ZWVBhd2RoME9xNVFrRXE3VG9KbGZjTWUySHFkT05STWJBRjFBWGdLWHdvazlyNWtadk95eE1EQjNQWHBNaTN0cDdtZmM4VVVCQmYzaFd2MjFNNUZabVdCWXV3?oc=5)
💡 À retenir
Weaviate released two patches (v1.38.5 and v1.36.22) with critical fixes for LSM store performance, batch vectorization deadlocks, and crash-safety, making them immediate priorities for production deployments. Anthropic's new mid-conversation system messages feature is now generally available for several of its top models, enabling more dynamic interactions.