Tag: Chine

Anthropic publie le 10 septembre un rapport nommant sept laboratoires chinois qui auraient distillé Claude à échelle industrielle. Le 22 septembre, Xiaomi — cité au dossier — met en ligne MiMo-V2.6 Pro, Flash et 9B en licence MIT, avec le meilleur score open-weight du marché. Coïncidence, provocation ou plan préparé ? Décryptage.

Le 14 septembre 2026, Chen Yixin — ministre de la Sûreté de l'État — publie dans China Cyberspace un texte long qui classe l'IA en « champ de bataille principal » de la compétition technologique et liste six menaces pour la Chine. Il cite nommément deux modèles américains. Au-dessus de lui, Chen Wenqing, patron de la Commission des affaires politiques et juridiques, valide la ligne. Décryptage.

Taiwanese researchers from TeamT5 document four Chinese groups — Grimfengxi, Teleboyi, Huapi, knaithe — using DeepSeek as an autonomous attack engine: 460 targets in 4 days, generated exploits, automated reconnaissance. The weakness of the Chinese model's safeguards is directly highlighted.

Alibaba unleashes Qwen 3.8 Max: 2.4 trillion parameters in total, 95 billion active, 1 million context tokens, native multimodal. The weights will be released open source one week after the announcement. Second largest open model in the world behind Kimi K3. China continues to flood the market.

Appearing anonymously on OpenRouter on August 20, OX Alpha outperforms GPT-5.6 Sol and Claude Fable 5 on code benchmarks — 80% on DeepSWE. Free until August 27, 1M context tokens, multimodal. Investigation into a fifth stealth model that clearly points to Z.AI.

Tencent Cloud releases the code for Team Memory, an extension of its TencentDB Agent Memory project (20,000 GitHub stars in 90 days) that allows multiple AI agents to share conversations, code, and documents with fine-grained governance. Industrial multi-agent systems are taking shape.

OpenAI has just made GPT-5.6 Luna its default free model for ChatGPT, following an 80% cut in API prices. The result: Luna now undercuts DeepSeek V4 Pro on input, increasing pressure on the Chinese price-to-quality ratio. Analysis of a strategic shift.

In June 2026, the US government ordered Anthropic to suspend access to Claude Mythos 5 for all foreign nationals. Then Alibaba banned Claude Code from its systems and released Qoder in response. Anatomy of an AI model that has become an instrument of foreign policy.

In a matter of weeks, three major Chinese models have been released in open-weights: DeepSeek V4 Flash (ultra-efficient LLM), MiniMax H3 (coherent video multimodal), and ByteDance Seedance 2.5 (AI video up to 7× cheaper than Sora and Veo). An overview and a reminder of what's at stake against Western players.

Moonshot AI released the open weights of Kimi K3, a 2.8 trillion parameter MoE, at the end of July 2026. A paid API exists in parallel. Breakdown: what's free, what's paid, and why the license isn't fully open.

Pékin s'oppose à l'acquisition par Meta de la startup d'agents IA Manus, valorisée plus de 2 milliards de dollars. Les fusions-acquisitions de l'IA deviennent un instrument de la rivalité sino-américaine.

Le 27 janvier 2025, un modèle chinois entraîné à prix cassé faisait perdre 600 milliards de dollars à Nvidia en une séance. Douze mois plus tard, bilan d'un électrochoc qui a redistribué les cartes de l'IA mondiale.