1/4 Following up on our launch of Tongyi DeepResearch: We're now releasing the full technical report! Dive deep into the technology and insights behind our 30B (A3B) open-source web agent that achieves SOTA performance: 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, and 46.7

JSON is token‑expensive for LLMs – just like @mattpocockuk frequently mentions. Meet TOON, the Token‑Oriented Object Notation. 💸 40–60% fewer tokens than JSON 📐 readable & tokenizer-aware Wrap your JSON with `encode` to save half the token cost: https://t.co/UoG9yHmgfg

Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other

🤯 Merging many finetuned LLMs into one model, effectively? Introducing Functional Dual Anchor (FDA), a new framework for model merging. 🚀 Current merging works poorly due to the underlying parameter conflicts. FDA shifts knowledge integration to the input-representation space

We, at NVIDIA, presents - Length Penalty Done Right - Cut CoT length by 3/4 without sacrificing accuracy using only RL - This makes DeepSeek-R1-7B running ~8 times faster on AIME-24 while maintaining the same accuracy.

👋Say Hi to MiMo-Audio! Our BREAKTHROUGH in general-purpose audio intelligence. 🎯 Scaling pretraining to 100M+ hours leads to EMERGENCE of few-shot generalization across diverse audio tasks! 🔥 Post-trained MiMo-Audio-7B-Instruct: • crushes benchmarks: SOTA on MMSU, MMAU,

On olmOCR-Bench, olmOCR 2 scores 82.4 points, up from 78.5 in our previous release—increasing performance across every document category. 📈

This is the JPEG moment for AI. Optical compression doesn't just make context cheaper. It makes AI memory architectures viable. Training data bottlenecks? Solved. - 200k pages/day on ONE GPU - 33M pages/day on 20 nodes - Every multimodal model is data-constrained. Not anymore.

This tool lets you scrape any documentation page and converts it into a Claude skill you can add to web, app, and Claude Code (Somebody should make a hosted version of this!) Lets see how links in the main post go on the new X… https://t.co/zaT2g5j5qE

Pretty wild what 900M params can do - PaddleOCR VL 🔥 > SOTA on OmniDocBench v1.0 & v1.5 (text, tables, formulas, charts, reading order) > Multilingual - 109 languages (Latin, Arabic, Cyrillic, Devanagari, Thai, etc) > Handles handwriting, historical docs, noisy scans > Supports

The open-source alternative to Durable Objects now runs on @vercel Functions ⚛️ Build collab apps, AI agents, and multiplayer games in Next.js 🔌 WebSockets on Vercel (you read that right) 📊 30x the memory (4GB vs 128MB) 💚 Standard Node.js runtime, no workerd

🚀 Introducing rLLM v0.2 - train arbitrary agentic programs with RL, with minimal code changes. Most RL training systems adopt the agent-environment abstraction. But what about complex workflows? Think solver-critique pairs collaborating, or planner agents orchestrating multiple

🚀 PaddleOCR-VL is here! Introducing PaddleOCR-VL (0.9B) — the ultra-compact Vision-Language model that reaches SOTA accuracy across text, tables, formulas, charts & handwriting. Breaking the limits of document parsing!🌍 Powered by: • NaViT dynamic vision encoder • ERNIE

@sundarpichai @Yale c2s-scale 27b by @david_van_dijk is live today btw hf: https://t.co/sxYw2mrBUa gh:

The paper shows that Group Relative Policy Optimization (GRPO) behaves like Direct Preference Optimization (DPO), so training on simple answer pairs works. Turns a complex GRPO setup into a simple pairwise recipe without losing quality. This cuts tokens, compute, and wall time,

🚀 The RL community keeps pushing boundaries — from better on-policy data and partial rollouts to in-flight weight updates that mix KV caches across models during inference. Continuing inference while weights change and KV states stay stale sounds wild — but that’s exactly what

🚀 Introducing 𝐀𝐠𝐞𝐧𝐭 𝐒3, the most advanced computer-use agent, now 𝐚𝐩𝐩𝐫𝐨𝐚𝐜𝐡𝐢𝐧𝐠 𝐡𝐮𝐦𝐚𝐧-𝐥𝐞𝐯𝐞𝐥 𝐩𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞🧠💻 Just one year ago, Agent S scored ~20% on OSWorld: SOTA then, but far from human 72%. Today, Agent S3 reaches 6̳9̳.̳9̳%̳ (⬆10% over

Announcing OpenMoE 2, the first-ever architectural study of sparse diffusion language models, trained from scratch. ✅ Expert-choice MoE × diffusion ✅ Ultra-wide FLOPs/param range (sparse → super-dense) ✅ Perfect load-balance (no aux loss) ✅ +20% throughput ✅ adaptive

🚿 SYS PROMPT LEAK 🚿 Here are the full sys instructions for Droid, the current top AI coding agent in the world! PROMPT: """ <Role> You are Droid, an AI software engineering agent built by Factory ( https://t.co/8mUTVYuBfV). You are the best engineer in the world. You write

🚀 New in vLLM: dots.ocr 🔥 A powerful multilingual OCR model from @xiaohongshu hi lab is now officially supported in vLLM! 📝 Single end-to-end parser for text, tables (HTML), formulas (LaTeX), and layouts (Markdown) 🌍 Supports 100 languages with robust performance on