🚀 Modèles & Releases
Anthropic has launched Claude Opus 4.8, its most capable generally available model, featuring a 1M token context window by default and 128k max output tokens. (https://docs.anthropic.com/en/release-notes/api#may-28-2026)
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)
Microsoft released two new embedding models on Hugging Face: microsoft/bitnet-embedding-0.6b and microsoft/bitnet-embedding-270m. (https://huggingface.co/microsoft/bitnet-embedding-0.6b)
Tencent released Hy-Embodied-VLM-1.0, a vision-language model for embodied AI and robotics. (https://huggingface.co/tencent/Hy-Embodied-VLM-1.0)
🔬 Recherche
A new paper introduces Trust Region Policy Distillation (TOP-D), a method to stabilize the notoriously unstable On-Policy Distillation (OPD) training paradigm. (https://huggingface.co/papers/2607.04751) A study explores leveraging unlabelled data to improve the generalizability of neural population decoders for brain-computer interfaces. (https://arxiv.org/abs/2607.14086v1) Research on Linear Independent Component Analysis via Optimal Transport presents a new method for recovering independent source signals from linear mixtures. (https://arxiv.org/abs/2607.14081v1) MetaPerch is a new approach for bioacoustics foundation models that learns from metadata to improve species detection. (https://arxiv.org/abs/2607.14072v1)
🏢 Industrie & Business
Applied Computing raised a $20M Series A to build a foundation AI model for the oil, gas, and petrochemical industry. (https://techcrunch.com/2026/07/15/applied-computing-wants-to-give-oil-and-gas-operators-an-ai-model-for-the-entire-plant/) Microsoft is reportedly training its salespeople to position its in-house AI models as more efficient and cost-effective than those from OpenAI and Anthropic. (https://techcrunch.com/2026/07/15/microsoft-is-reportedly-training-salespeople-to-talk-down-openai-and-anthropic/) Amid a legal battle with Apple over hardware trade theft allegations, OpenAI has released a $230 light-up keyboard designed for its agentic coding app, Codex. (https://techcrunch.com/2026/07/15/amid-hardware-legal-battle-openai-releases-a-230-keyboard-for-codex/)
🛠️ Outils & Produits
Ollama v0.32.1 improves Gemma 4 tool calling and multi-turn reasoning, fixes an MLX model cache leak, and improves agent web search. (https://github.com/ollama/ollama/releases/tag/v0.32.1-rc0) AutoGPT Platform v0.6.67 adds first-class organization/workspace support and splits the adapter base into Socket/Webhook components. (https://github.com/Significant-Gravitas/AutoGPT/releases/tag/autogpt-platform-beta-v0.6.67) Cline CLI v3.0.42 fixes Ollama native API routing so context window and timeout settings work again. (https://github.com/cline/cline/releases/tag/cli-v3.0.42) Crush v0.85.0 introduces a first-class question tool for LLMs, supporting single choice, multiple choice, and free-form text inputs. (https://github.com/charmbracelet/crush/releases/tag/v0.85.0)
⚙️ API & Dépréciations
Anthropic has released a beta Admin API for Claude Enterprise organizations, enabling programmatic management of members, roles, groups, and invites. (https://docs.anthropic.com/en/release-notes/api#july-14-2026) NVIDIA TensorRT-LLM v1.3.0rc21 announces the deprecation of the AutoDeploy backend, with the team working on agentic approaches for faster model support in the PyTorch backend. (https://github.com/NVIDIA/TensorRT-LLM/releases/tag/v1.3.0rc21) Weaviate released patches v1.36.22 and v1.38.4, focusing on LSM segment cleanup, WAL recovery crash-safety fixes, search-time HNSW entrypoint repair, and async replication performance. (https://github.com/weaviate/weaviate/releases/tag/v1.36.22) Llama.cpp commit b10032 adds a CUDA implementation for the GGML_OP_LIGHTNING_INDEXER operation. (https://github.com/ggml-org/llama.cpp/releases/tag/b10032)
⚖️ Régulation & Société
Japan has revised its AI policy guidelines to bolster cybersecurity measures. (https://news.google.com/rss/articles/CBMilAFBVV95cUxNU3U5ZHNqVjJxb0EwTWk2VjJIN2xDelhySXhKTTJBMDNqU01mWWpXYkVmU1NLUmJCemNhTnpKTDBJZTNqTmtSOWIyX0tGZ1RFZWZhRmpsUFFCb1QxVG50TlNwTlBfUTV4bW81OGtHY2ZKTDQtVFZVX3I4T1o2RjY5TklCUU51UHFiZW1WemloRFV3STZ6?oc=5) A new analysis provides a global comparison of AI regulation frameworks. (https://news.google.com/rss/articles/CBMieEFVX3lxTE1RQ2FWUGd2bVhld29vV29tVFR3NjBPMWxPVVFmeFJTRnNYSG43VnlGd3BZdmxuOTR3NUV6eXlkZHN4SElNTk5WSkQ5WkY3aWdoWmY0UjFCWHZ0U1FQdUtDd29ITHVYXzdGSmJuM1NhVjZPeEhlNXBhag?oc=5) An op-ed by Rep. Max Rose argues that Congress should look to the successes of the Internet Age when crafting AI regulation. (https://news.google.com/rss/articles/CBMi1wFBVV95cUxPaGswLXBONTVCVWpib2dGempQaHh4aXhLbUF2czlJRFduU3pQUWRqaFBER3VWQVBzeWhZRUZtLUt4c1E3Sk5IOVNaXzZoajQxVXRncWprR1pxOXhhbTg5b2VLWkNEUUdEUmV5YlpTRXRrRGoxWkN2d1VJRS1xQi1UODVHTXZnZ3FmY1ByRi14aWJtd1R0YjZXVVM3ZWNwWENJN3F6R1htVXE3anlrU2RPSjZfRll5S2dDazlxS0g3bWV5d1JlZmNtYzB2OHZCUnY5UlhsTHpCTQ?oc=5) OpenAI has built an LLM super-hacker called GPT-Red, which it uses as a sparring partner to improve the cybersecurity defenses of its other models like GPT-5.6. (https://www.technologyreview.com/2026/07/15/1140514/meet-gpt-red-an-llm-super-hacker-openai-built-to-make-its-models-safer/)
💡 À retenir
The most critical update is Anthropic's new beta Admin API for Claude Enterprise, which allows backend teams to programmatically manage organization members, roles, and groups—automating a key administrative workflow.