Zac Zuo

Hi everyone！Sharing SmolVLM2, a new open-source multimodal model series from Hugging Face that's surprisingly small, with the smallest version at only 256M parameters! It's designed specifically for video understanding, opening up interesting possibilities for on-device AI.What's cool about it:📹 Video Understanding: Designed specifically for analyzing video content, not just images.🤏 Tiny Size:...

Hi everyone!Check out something very different in the LLM space: Mercury, from Inception Labs. This isn't another autoregressive model like GPT or Llama, it's a diffusion LLM (dLLM), and it's the first commercial-scale one.The big deal here is speed. Diffusion models are how AI generates images (think Midjourney, Stable Diffusion), but applying them to text is new. Mercury is up to 10x faster...

Hi everyone!Sharing DeepGEMM, a new open-source library from DeepSeek. This is definitely for the hardcore GPU optimization crowd, but it's a great example of how much performance can be squeezed out of specialized hardware.DeepGEMM provides highly optimized FP8 GEMM (General Matrix Multiplication) kernels specifically for NVIDIA Hopper GPUs. It's designed to power both dense models and...

Hi everyone!Sharing Magma, a new open-source foundation model from Microsoft Research, and it's a big deal for AI agents. Unlike most multimodal models, Magma isn't just for understanding images and text, it's designed to act in both digital and physical environments (think UI navigation and robot manipulation).It's all based on a new pretraining approach using Set-of-Mark (SoM) and...

Hi everyone! PixVerse V4 is here, and it's a major upgrade to the PixVerse AI video generation platform! They've focused on more realistic videos, faster generation, and added some completely new features. Key improvements: ✨ Enhanced Realism: Videos should look more realistic overall. 🏃 Faster Generation: Speed improvements across the board. 🔊 Sound & Speech Generation: This is HUGE!...

Hi everyone!Google launched a free version of Gemini Code Assist for individual developers! This is the same AI coding assistant they've been offering to businesses, now available to everyone.What's the deal?💻 IDE Integration: Works directly inside VS Code and JetBrains IDEs.🤖 Gemini 2.0 Power: Powered by a coding-optimized version of Gemini 2.0.💬 Code Completion, Generation, & Chat: Get...

Hi everyone！<think>...</think> is cool, especially with Qwen.QwQ-Max, a new LLM from the Qwen team and it's focused on deep reasoning!What's interesting about this:🧠 Deep Reasoning: It's built on Qwen2.5-Max and specifically designed for complex reasoning, math, coding, and agent tasks.🤔 "Thinking Mode": You can activate a special "thinking mode" with a system prompt ( tags) to...

Hi everyone!Sharing DeepEP, a new open-source communication library from DeepSeek AI, and this is the technology that powers their impressive DeepSeek-V3 model! It's a highly technical project, but crucial for anyone working with Mixture-of-Experts (MoE) models at scale.DeepEP is specifically designed to optimize the all-to-all communication that's critical for MoE training and inference. It's...

Hi everyone! Dancing with the Moonlight? Well, it's a new family of open-source language models from Moonshot AI (the creator of Kimi.ai) that's pushing the boundaries of LLM training efficiency. The key here is the Muon optimizer, it achieves comparable performance to AdamW-trained models with only half the compute! What's interesting: 🤖 3B and 16B MoE Models: Available in two sizes, both...

Hi everyone! It's interesting to see Apple expanding News+ beyond just news and magazines. It seems like they're aiming to make it a more comprehensive content hub. Apple News+ Food is a dedicated section for all things food, and it's being added to the existing Apple News+ subscription (which is $12.99/month in the US). Certain content will also be available to non-subscribers. The...

Hi everyone！Aider is an open-source tool that brings AI pair programming directly into your terminal. It lets you collaboratively edit your existing codebase with an LLM, right in your terminal.Here's how it works:Add Files: You tell aider which files to work on, like: aider file.jsChat: Describe changes in natural language.AI Edits: aider applies edits directly to your files.Review &...

Hi everyone! Sharing ToddlerBot, an open-source humanoid robot platform from Stanford. It's designed to make advanced robotics research more accessible. What's particularly interesting is that it's built for both locomotion and manipulation, and it's surprisingly low-cost(less than 6K USD). Key features: 🤖 30-DOF humanoid design (up to 43 DOF). 🚶 Designed for both walking and...

Good one! The Name is Winning :)