Les Émirats arabes unis viennent de lâcher K2-Think, le modèle open source de raisonnement le plus rapide du monde, et ça pulvérise tout ce qu’on croyait savoir sur le scaling de l’IA. 32 milliards de paramètres. C’est tout. Et ce truc atteint le niveau de raisonnement de GPT-4
OPEN SOURCE
-
2025 LLMs Compared: Architecture Differences and Technical Innovations
By
–
Comparing & Contrasting Recent LLMs Architecture > DeepSeek-V3/R1
> OLMo 2
> Gemma 3
> Mistral Small 3.1
> Llama 4
> Qwen3 (dense+MoE)
> SmolLM3
> Kimi 2
> GPT-OSS Are 2025 LLMs really that different from each other? MoE, MLA, GQA, sliding window, normalization games & more. -
HuMo: ByteDance’s Open-Source Multimodal AI for Video
By
–
Fast, high-res videos up to 720p@25fps! HuMo’s flexible, multimodal AI is built for creators & researchers pushing boundaries. Get into the future of video creation—open-source and ready to explore! #ByteDance #AI #HuMo #OpenSource
-
HuMo: Open-Source Video AI for Character Synthesis
By
–
🎵 Sync audio perfectly with video and preserve subject identity throughout your clips.
— Futurepedia – Learn to Leverage AI (@futurepedia_io) 13 septembre 2025
HuMo lets you bring characters to life with incredible detail and consistency.
Imagine the possibilities! 👀 #OpenSource #VideoAI #AIInnovation pic.twitter.com/q0g0HIWiKPSync audio perfectly with video and preserve subject identity throughout your clips. HuMo lets you bring characters to life with incredible detail and consistency. Imagine the possibilities! #OpenSource #VideoAI #AIInnovation
-
100+ Free Open Source AI Agents and RAG Tutorials
By
–
Stay tuned for more such interesting posts → @Saboo_Shubham_ I have created 100+ AI Agents and RAG tutorials, 100% free and opensource. P.S: Don't forget to star the repo to show your support
-
VLA-Adapter: Effective Paradigm for Tiny-Scale Vision-Language-Action Models
By
–
VLA-Adapter An Effective Paradigm for Tiny-Scale Vision-Language-Action Model
-
VaultGemma: Open Model with Differential Privacy Scaling Laws
By
–
VaultGemma is a release of an open model trained from scratch with differential privacy. The blog post below and the full tech report linked from the tech report have some nice analyses to present a scaling law for differentially private language models: Blog:
-
Meta Releases MobileLLM-R1: Compact Edge Reasoning Model Under 1B Parameters
By
–
Meta just dropped MobileLLM-R1 on Hugging Face a edge reasoning model with fewer than 1B parameters 2×–5× Performance Boost over other fully open-source models: MobileLLM-R1 achieves ~5× higher MATH accuracy vs. Olmo-1.24B, and ~2× vs. SmolLM2-1.7B. Uses just 1/10 the
-
Genspark’s Free AI Browser with Local Models
By
–
Genspark introduced a free AI browser that works with local models! You can choose from 169 open-weight models to keep your conversations private.
— 🚨 AI News | TestingCatalog (@testingcatalog) 12 septembre 2025
It also comes with built-in Super Agents 👀 pic.twitter.com/040PXUR2IpGenspark introduced a free AI browser that works with local models! You can choose from 169 open-weight models to keep your conversations private. It also comes with built-in Super Agents
-
HuggingFace redefines identity as AI builder journey starting point
By
–
This feels like HuggingFace doubling down on their identity: Not just a model hub… but the place where the AI builder journey starts.