🚀 Modèles & Releases
Thinking Machines Lab released Inkling, an open-weight multimodal mixture-of-experts model with 41B active parameters out of 975B total, designed for general-purpose reasoning, coding, and agentic systems. (https://openrouter.ai/models/thinkingmachines/inkling) Moonshot AI released Kimi K3, a 2.8T parameter open-weight multimodal reasoning model suited for complex coding, knowledge work, and long-horizon agentic workflows. (https://openrouter.ai/models/moonshotai/kimi-k3) Meta released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, accepting text, images, video, audio, and PDF documents with a 1M-token context. (https://openrouter.ai/models/meta/muse-spark-1.1) OpenRouter launched Auto Router (Beta), a task-aware router that classifies requests and routes them to the most popular model for that task based on aggregate spend. (https://openrouter.ai/models/openrouter/auto-beta) Kwaipilot 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) InternLM released Intern-S2-Preview-397B and an FP8 quantized version, Intern-S2-Preview-397B-FP8, as image-text-to-text conversational models. (https://huggingface.co/internlm/Intern-S2-Preview-397B, https://huggingface.co/internlm/Intern-S2-Preview-397B-FP8) Microsoft released microsoft/bitnet-embedding-0.6b, a multilingual embedding model. (https://huggingface.co/microsoft/bitnet-embedding-0.6b)
🔬 Recherche
A paper titled "Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models" explores in-context learning as conditional inference and the probabilistic consistency of LLM outputs. (https://arxiv.org/abs/2607.15277v1) "RoboTTT: Context Scaling for Robot Policies" introduces a method to scale visuomotor context to 8K timesteps, three orders of magnitude beyond current state-of-the-art policies. (https://arxiv.org/abs/2607.15275v1) "MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators" proposes a reinforcement learning method to align efficient flow-based generators with human preferences. (https://arxiv.org/abs/2607.15273v1)
🏢 Industrie & Business
Databricks has reached a $188B valuation, extending its run as an AI-focused company and publishing research on the cost savings of open-weight AI models for coding. (https://techcrunch.com/2026/07/17/databricks-hits-188b-valuation-extending-its-run-as-ais-favorite-second-act/) Vertu is marketing a $6,880 luxury foldable phone with an integrated AI agent, targeting executives. (https://techcrunch.com/2026/07/17/vertu-wants-executives-to-pay-6880-for-an-ai-agent-heres-how-it-actually-performs/)
🛠️ Outils & Produits
Google announced updates to Search allowing users to connect more apps for rendering and to Google Vids for creating, editing, and starring in videos with Gemini Omni and Personal Avatars. (https://blog.google/products-and-platforms/products/search/connected-apps/, https://blog.google/products-and-platforms/products/workspace/gemini-omni-personal-avatars/) CrewAI released version 1.15.3, adding features like organization ID parameters, step interception points, and declarative flow execution on the TUI. (https://github.com/crewAIInc/crewAI/releases/tag/1.15.3) Llama.cpp commit b10052 includes a Hexagon L2 cache handling rework and more MUL_MAT updates for performance. (https://github.com/ggml-org/llama.cpp/releases/tag/b10052) Microsoft GraphRAG released v3.1.1, adding a native CosmosTableProvider with namespace partitioning and updating the LiteLLM dependency. (https://github.com/microsoft/graphrag/releases/tag/v3.1.1)
⚙️ API & Dépréciations
Weaviate released v1.37.13 with performance fixes for BM25 and LSM reads, and search correctness fixes. (https://github.com/weaviate/weaviate/releases/tag/v1.37.13)
Weaviate released v1.38.5 with LSM store performance improvements, a fix for potential batch vectorization deadlocks, and allocation-free reference cache improvements. (https://github.com/weaviate/weaviate/releases/tag/v1.38.5)
Anthropic made mid-conversation system messages available on Claude Fable 5, Claude Mythos 5, and Claude Opus 4.8 via its API, Amazon Bedrock, and Google Cloud, with no beta header required. (https://docs.anthropic.com/en/release-notes/api#july-15-2026)
The OpenAI Python SDK was updated to v2.46.0, adding an endpoint for managing organization project service account API keys. (https://github.com/openai/openai-python/releases/tag/v2.46.0)
Pydantic-AI released v2.13.0, adding features like an include_model_request_parameters instrumentation setting and a RaiseContentFilterError capability. (https://github.com/pydantic/pydantic-ai/releases/tag/v2.13.0)
⚖️ Régulation & Société
The EU AI Act's first major provisions are set to apply on August 2, 2026, prompting analysis on what actually applies and how employers should prepare for workplace AI. (https://news.google.com/rss/articles/CBMikAFBVV95cUxQQlFIdnRpdVdDbzEzWmo1ZjhhTFdyQXZRX29OOTZSdGhzQ25JZm8zUFdQanlEQjg5Z0N3d2hrU3U4Vkt2QkNBYzVXbG93Q3N5OWc2ZG4zaWM1dGNfQW9GYXp5dVhkZlA2QUVRUG1DOEt6d3ZzNTh0VnQzNmFwaGl0NnhWU2J6WjJpblVWMkswTEY?oc=5, https://news.google.com/rss/articles/CBMixwFBVV95cUxPMUVfTXRUaE1zQ3EybU92RTZoQlBNWEJmNFYtZDlneDNyZTVFX2dWUmJRTjF2SmdQTTNYOU0xWmpRaVFSRmRJU2N1N3hXNE4wSFQtSy1YaTZkbmw4Mkd1XzI1MkVJcTQ4bDNSZkQyZWNsV1htZHpRVUI2M0R4cmQ2Y1daaUFLSkhfSU9wZlU3SHdrRnRtZkd6QkV0YVVHVEsyVmFWNGFnU09oWDZTeHlBWi1kWTEyMVFTQUlHdW5mdHZGVHRwTDhj?oc=5) Employees at Anthropic have donated $3M to support AI safety regulations. (https://news.google.com/rss/articles/CBMiekFVX3lxTE4tbUpnNm9ncDhWME9GbFc3SVBzTWJnLVBYNzllWGRsLU9XbmpwTTkxdWVWVUpHRnR4dDVXeFp6ZFNUNS1GZG4zQjQwMmxUSlRsM21aR1dhVmx3TlplNmJYdDAydHdKMlB1WERSQzkybkFKNGFFR3ZGMFJ3?oc=5) A report highlights the rising risk of weather data sabotage, noting its critical influence on major strategic decisions across industries. (https://www.technologyreview.com/2026/07/17/1140622/weather-data-sabotage/) An article discusses the privacy implications of ubiquitous meeting transcription and summarization, exemplified by a "Don't record me" Zoom hack. (https://techcrunch.com/2026/07/17/the-zoom-hack-that-says-dont-record-me/)
💡 À retenir
Weaviate released two patches (v1.37.13 and v1.38.5) with critical performance and correctness fixes for production deployments, making them the top priority for immediate updates.