Deux sujets techniques : petite analyse de Claude Code, puis montage d’une machine d’AI coding isolée en 30 minutes chrono — Proxmox, VM, réseau cloisonné, environnement prêt à coder ou l’inférence.

meetup https://www.meetup.com/grenoble-open-artificial-intelligence-meetup-group/events/314150636/?slug=grenoble-open-artificial-intelligence-meetup-group&eventId=314150636

Une actualité chargée avec de nombreuses avancées : éditeurs d’images physiques, avatars VR temps réel, génération de polices vectorielles, modèles Qwen optimisés, robots humanoïdes, et bien plus.

• Mistral revient dans la course avec Devstral 2
• Tendance 2025 et 2026 sur le code et les agents
• Analyse de “Automating Massive Parallel AI Coding”, présentation d’OpenHands (partenaire de Mistral AI)
• Anthropic lance Claude Opus 4.5, un nouveau modèle surpuissant pour le code
• Nouvelles tendances sur le développement assisté par IA

Mistral

Mistral annonce Devstral 2 montrant une stratégie payante sur l’industrialisation de l’écriture de code. Devstral 2 offre des modifications fill-in-the-middle dans un répertoire de code basé sur une instruction en langage naturel. Rivalisant avec les modèles frontiers comme Sonnet 4.5, pas encore Opus 4.5.

Sur OpenRouter, Devstral 2 a une version payante et une version gratuite mistralai/devstral-2512:free (dépréciée le 27 janvier 2026).

“Devstral 2 is a state-of-the-art open-source model by Mistral AI specializing in agentic coding. It is a 123B-parameter dense transformer model supporting a 256K context window. Devstral 2 supports exploring codebases and orchestrating changes across multiple files while maintaining architecture-level context. It tracks framework dependencies, detects failures, and retries with corrections—solving challenges like bug fixing and modernizing legacy systems. The model can be fine-tuned to prioritize specific languages or optimize for large enterprise codebases. It is available under a modified MIT license.”

Mistral annonce devstral2 montrant une stratégie payante sur l’industrialisation de l’écriture de code. Devstral2 offre des modifications fill-in-the-middle dans un répertoire de code basé sur une instruction en language naturel. Rivalisant avec les modèles frontiers comme Sonnet4.5, pas encore Opus4.5. Sur OpenRouter Devstral2 a une version payante et une version gratuite mistralai/devstral-2512:free (Deprecating 27 janvier 2026)

“Devstral 2 is a state-of-the-art open-source model by Mistral AI specializing in agentic coding. It is a 123B-parameter dense transformer model supporting a 256K context window. Devstral 2 supports exploring codebases and orchestrating changes across multiple files while maintaining architecture-level context. It tracks framework dependencies, detects failures, and retries with corrections—solving challenges like bug fixing and modernizing legacy systems. The model can be fine-tuned to prioritize specific languages or optimize for large enterprise codebases. It is available under a modified MIT license.”

« Mistral AI, le champion français de l’IA générative, officialise une levée de fonds de 1,7 milliard d’euros. Au tour de table, on retrouve ASML, géant néerlandais des puces électroniques, qui injecte 1,3 milliard d’euros. Mistral AI est désormais valorisée à 11,7 milliards d’euros. […] Ce qui en fera le premier actionnaire de la start-up après les fondateurs, qui contrôlent encore majoritairement la société. »

Partage de quelques nouvelles sur les avancées concernant les modèles, agents, local-first, RAG et capacités multimodales. Cette mise à jour propose un tour d’horizon des derniers progrès — de ChatGPT-5 et GPT-OSS20b/120b aux modèles « banana » de Google, en passant par les frameworks open-weight comme Qwen3, ainsi que des agents de codage.

L’intelligence artificielle (IA) a franchi en 2025 un nouveau cap de maturité, avec des avancées majeures sur toutes ses modalités fondamentales : texte, image, vidéo, audio et 3D.

The AI landscape continues to evolve rapidly with two major developments this month: the release of Qwen3 and OpenAI’s strategic acquisition of Windsurf. These developments highlight the intensifying competition in both the general AI and coding assistant markets.

AI can be extremely useful for searching and sorting through large amounts of information on the internet. One of its most natural uses is to search for and extract sources. The internet allows us to make connections between multiple pages, which makes recursion a key issue. In modern AI tools such as ChatGPT, OpenAI has chosen to make their “advanced search” service available. Although quite basic in reality, this service enables an AI “agent” (an agent is a program that acts autonomously to accomplish a specific task) to search the web and extract relevant information from the pages it visits. This search allows access to high-quality, relevant content for a given topic and research area. This makes it an excellent source of information for generating prompts later on!

AI as a General-purpose technology to understand what is going on

General-purpose technologies (GPTs) are technologies that can affect an entire economy (usually at a national or global level).GPTs have the potential to drastically alter societies through their impact on pre-existing economic and social structures. The archetypal examples of GPTs are the steam engine, electricity, and information technology.

If you missed recent AI newsk, a new open source model was released on 20th of january, DeepSeek-R1 from Chinese researchers.

If you missed recent AI news and tools for creators, look at Comfyui. The Comfyui team is expanding fast as a local-first AI tool for a broad range of AI generation tasks. Mainly oriented on image and video, it is also very powerfull to be adaptable to almost all uses of AI.

New release for 3D mesh generation. A LLM used as the main building block for 3D mesh generation. Quicktip: Compare AI tools: https://www.blogdumoderateur.com

This update covers the latest advancements in open-source AI interfaces for chatbot conversations, image generation, tooling, and more.

Mistral the French startup that release at an impressive pace new models, keep delivering on local AI ! Today a new magnet link was published to introduce Pixtral, a new multimodal vision model.

https://x.com/mistralai/status/1833758285167722836?s=46

Companies developing proprietary AI models are finding themselves in a constant sprint to keep pace with the rapid advancements in open-source alternatives.

If you’re reading this, chances are you’re planning to build a local AI system with a single RTX 4090. But before you dive in, let me share some valuable insights that can help you save a significant amount of money while achieving your goals.

As AI services proliferate, companies and individuals are faced with an increasing complexity of providers, microservices, and terms of use.