🚀 NVIDIA just dropped the biggest CUDA update in 20 years. 👌 CUDA Tile lets you write GPU code without managing threads. 🔥 No more threadIdx calculations. No manual memory juggling. ✌️ I broke it down in 10 minutes with actual code examples 👇 https://t.co/mIm3wfxEow

🚀 GAME-CHANGER ALERT! VibeVoice-Realtime makes AI talk INSTANTLY! 🤯 Just 300ms latency & LLMs can speak from their FIRST tokens! No more waiting for complete responses. 🔊 Streaming text input ✅ 🎤 Single speaker focus ✅ ⚡ Lightning-fast response ✅ Watch video now:

Apple Surprises with CLaRa-7B: A Useful RAG Model 🚀Breaking: Apple's CLaRa AI revolutionizes RAG! 🔥Compresses docs into smart, retrievable embeddings via QA/paraphrase training, ♨️then jointly optimizes retrieval & generation in one continuous space. Full Local Demo:👇

Please don't treat AI models like Crypto coins.

@fahdmirza this one is reasoning, I haven't tested but super curious

Meta AI's SAM 3D is Live on fal! You can turn a single image file into an actual 3D Object! ✅ 3D Alignment ✅ Image to 3D (Human) Body ✅ Image to 3D Object Link 👇🏽

🎉 Meet Ministral 3 3B Instruct! A tiny, open-source model designed for the edge. ♨️ Vision: Analyze images & text ♨️ Local: Runs efficiently in ~8GB VRAM ♨️ Multilingual: Supports dozens of languages ♨️ Agentic: Native function calling & JSON output ♨️ Large Context: 256k

This might just be my favorite release of the year, especially with everything under Apache 2.0! 🔥 Which model are you most excited about?

@fahdmirza @MistralAI Thanks for the video! 🙌

It's here! 🚀 Introducing Mistral Large 3, most powerful model yet. A state-of-the-art multimodal MoE with 675B total parameters, 256k context, and frontier-level performance. Best of all? It's open-source under Apache 2.0. Dive in 👉 https://t.co/ikldUUTWD9 @MistralAI

Transformers v5: Simple model definitions powering the AI ecosystem @huggingface

The vision: transformers as the backbone for the entire open AI/ML stack - training, fine-tuning, inference. The future is here. 🚀 This format makes the key metrics stand out while keeping it concise and impactful.

Built for scale: • PyTorch-native architecture • Modular design patterns • Quantization-first approach • OpenAI-compatible API with Responses support

Transformers v5 is here: The Numbers Don't Lie 📊 🔥 Daily Installs: 20k → 3M+ (150x growth!) 🏗️ Architectures: 40 → 400+ supported models 📦 Checkpoints: ~1k → 750k+ available 🎯 Total Installs: 1.2B+ across all versions

Token Intelligence and MoE routing is exceptional in this new version.

https://t.co/zIcrtGocJj

🚀 DeepSeek-V3.2 just dropped: ✨ Matches GPT-5, Speciale beats it 🥇 Gold medals: IMO, IOI, ICPC 2025 ⚡ 60% cheaper on long context (128K tokens) 🤖 Best open-source coding agent (73% SWE-Verified) 🧠 Thinks WHILE using tools (game changer) 💰 10%+ of pre-training spent on RL

AI has to rise above Attention, if it's to become AGI

For me the, The Release of Last week was: Z-Image model from @Ali_TongyiLab

As thanksgiving gift @Kimi_Moonshot has given limited free access: https://t.co/4ToMhqQMOZ