I'm going to try the new @NVIDIAAI Nemotron-3-Nano-30B-A3B and compare it to Qwen 3.6 35B in agentic workflows. https://t.co/z9cnRBOo1c

🛠️ Agent Customization Customize AI workflows with agents, instructions, skills, prompts, and hooks. 🔗 https://t.co/ag5zffSLjd https://t.co/NSP4H9DwYj

Take Fable 5 for a spin in Cursor:

Introducing Devin Security Swarm A more cost effective and accurate way to find security vulnerabilities in complex codebases, based on a new architecture: Agentic MapReduce.

📢 1) We have a few papers that advance the state of the art of AI agent evaluation. Details and links in Stephan's post. 2) AI agent evaluation has quickly become a distinct discipline. We're working on a paper titled "Emerging trends in AI agent evaluation" that extracts best practices for this community. 3) I'm giving an invited talk at ICML, addressing anxiety about supposedly imminent Recursive Self Improvement and the question of what will remain for humans to work on (especially scientists, researchers, software engineers). I hope to make it provocative but cautiously optimistic. https://t.co/rYHlxPGEXY (I also plan to share the ideas from the talk as essays on the AI as Normal Technology newsletter.)

It is worth being very, very careful about how you are approaching routing, especially when the systems are primarily tested on verifiable IT benchmarks, which may lead you to overestimate the ability of weaker models.

Deeper Instructions, Stronger Generalization: Training on ComplexConstraints Given the chance, a model will reward hack however it can: finding the laziest path that satisfies a grader, whether or not that path reflects what you actually wanted. If the grader can be satisfied by a surface trick, that trick is what the model learns. Most instruction-following benchmarks are full of surface tricks. "Stay under 300 words," "avoid commas", a model can satisfy those by scanning the output text, without understanding the task at all. ComplexConstraints, our frontier instruction-following benchmark, is built so there's no lazy path: its constraints fire only under certain conditions, depend on the outputs of earlier steps, require planning ahead, and are often left unstated. You can't satisfy "don't assign anyone with a religious dietary restriction to pork prep" by pattern-matching. You have to understand who's who and reason through many interdependent requirements at once. We post-trained Qwen3-4B on 1,000 of these tasks, using expert-written rubrics directly as the RL reward. The results: → +15.5pp on the held-out set, reaching parity with a model 60x larger → the gains transferred to two external benchmarks the model never trained on: +8.4pp on Meta's AdvancedIF and +10.1pp on MultiChallenge → the largest gains landed on multi-turn abilities, even though every training example was single-turn Think about that last result. When the only way to score is to actually track many interdependent requirements, the model learns that skill rather than a shortcut, and the skill is the same whether the requirements arrive in one complex prompt or accumulate over nine turns. So it showed up on tasks the model was never trained on. A reward signal is only as good as the thought behind it, and not all rubrics are created the same. Research Blog: https://t.co/bUJPcoNFrX Research Paper: https://t.co/zQxE0TN260

"That is the difference between using a coding agent and engineering an autonomous coding system. One gives you a conversation. The other gives you a harness." https://t.co/47NWbraF3G < I liked the descriptions and visuals from @omarsar0 here. Very understandable! https://t.co/nIthf99EMB

One runtime, multiple GPU architectures, and zero vendor-specific model code. In this blog post, the TokenSpeed team @lightseekorg introduces TokenSpeed-Kernel, a portable, high-performance kernel system built for modern LLM inference. Using GPT-OSS 120B as a case study, they show how specialized kernels for @AIatAMD and @NVIDIAAI GPUs can seamlessly coexist behind a common API. This unified approach delivers up to 3.6x higher throughput on the AMD MI355X, all without requiring any changes to the underlying model logic. Link to blog in comments section 👇

Deploying AI models at the edge comes with a different set of challenges. These hands-on Jupyter labs walk you through using ExecuTorch to deploy and optimize @PyTorch models on Arm CPUs and NPUs, with examples you can run on hardware including Raspberry Pi. https://t.co/mJv4hbYFUZ

Most AI audio models have never heard a maqam. Team Motif fine-tuned Stable Audio 3.0 on Arabic maqam, built an Ableton plugin for microtonal style transfer, and won our Stable Audio 3.0 Challenge at Music Hackspace running locally on device. Watch Jad Al Masri break it down 👇

The Waypoint-1.5 technical paper is now live. Waypoint-1.5 is a real-time video diffusion world model designed to run on consumer GPUs, bringing interactive world models closer to practical, accessible deployment. https://t.co/U04x1YEwhF

.@tufalabs just open sourced their 1st place notebook 👀 https://t.co/tLs8aNmJ7P

Sakana Fugu Technical Report Instead of training one larger model, Sakana AI trains an orchestrator that reads each query and dynamically routes or composes GPT-5.5, Gemini-3.1-Pro, Claude Opus 4.8 and other agents into query-specific workflows. With Fugu being the fast router, and Fugu-Ultra being the deep multi-agent conductor, trained with SFT, evolutionary strategies and GRPO to build adaptive scaffolds. The idea is to have the model pick GPT for math, Gemini for science and recall, Opus for debugging, then synthesize them when no single agent is best. This router is able to get SoTA results across SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, CharXiv and more, demonstrating the potential of orchestration being a practical alternative beyond training.

Announcing the first production robot navigation framework on $500 hardware Explore the world once → your robot agent will relocalize and build a persistant, spatial memory across sessions SLAM, relocalization, loop closure, map i/o, planning, control No ROS. Open source. https://t.co/VCk9GvOrrM

We’re introducing GeneBench-Pro, a research-level benchmark for a harder kind of AI progress: how well agents can navigate messy biological data, choose the right analysis path, and make judgment calls that real computational research depends on. https://t.co/AsilnnSxnE

@Etched Congrats!! I was impressed to learn about some of the engineering wizardry (e.g. *very* low voltage domains, cluster scale memory, ...) that goes into tokens/watt maxxing of state of the art LLMs at interactive tokens/sec/user. Esp fun and memorable is the idea that this is engineering at the "opposite" regime to that of power transmission lines: very low voltage high current (at tiny distances) vs. very high voltage & low current (at great distances). Looking forward to more!

If you ever wondered about how how open/closed model makers and inference providers make economic sense, this is the piece to read

New benchmark added to Papers with Code based on @giffmana's Schmidhubering 🫡 Check the SOTA for semi-supervised ImageNet (using 10% of the labels) here https://t.co/CXd4lLkhlG https://t.co/sGi68AIoqh

HalluHard update: We’ve added GLM-5.2, using adaptive thinking with maximum reasoning effort, to our leaderboard. Despite its impressive performance on other benchmarks, GLM-5.2 still hallucinates frequently on our challenging multiturn benchmark. https://t.co/xbppFeo7Pd

DeepSeek preparing release of DSpark, DFlash and Eagle draft models for Qwen3 and Gemma-4 variants https://t.co/2zdfL9XAkQ

Got the model converted to CoreML and working on iOS; will open source soon! https://t.co/6xo8VetVGT

Today, we give robots a /skills library that self-evolves and compounds indefinitely! Introducing ASPIRE: a robot solving its 100th task is no longer as clueless as solving its first. Coding agents observe multimodal sensory traces from simulation and real robots, launch an evolutionary search over control programs, and distill the best know-how into an ever-expanding library. ASPIRE is a new type of continual learning: "training" is skill refinement instead of gradient descent. "Trained model" is a repo of sensorimotor skills instead of floating weights. “Distributed training” is a panel of agents each practicing a different skill instead of sharded minibatches. Here's the beauty: ASPIRE gives the tired terms "sim2real transfer" and "cross-embodiment transfer" a whole new meaning. Bridging the sim-to-real gap is notoriously brutal. An end-to-end policy has to swallow both the visual shift (sim looks toyish next to a real camera) and the subtle contact physics it never quite gets right. ASPIRE sidesteps the mess, because it doesn't ship pixels or weights across the gap, but ships the know-how. The robot still has to practice in the real world, not zero-shot, but it gets there way faster because it isn't rediscovering the strategy from scratch. Same for going single-arm to bimanual hardware, which usually requires new data and retraining from zero. ASPIRE achieves up to ~10x cut in "transfer learning” tokens (yes, tokens are the new unit of *training* compute ;) Check out our gallery of 150+ tasks and 90+ skills the robots taught themselves, all on the website! Kind of wild that we can ship the "learned weights" as an HTML page rather than a GGUF. We'll open-source the full stack so your own robot library starts compounding from ours! Deep dive in thread:

Our team at Xaira was fortunate to have early access to test Claude Science (Operon). 🔥🚀 We used it to add agentic loops to both virtual cell modeling and protein design workflows. A nice plus: Operon had already added our scGPT as one of the default skills for single-cell analysis 🙏😎🔥 This is the kind of product that actually understands how research works, not just chat with a model, but traceable artifacts, reproducible environments, and real scientific data connections. That's a big deal for computational biology.

Can regularization based JEPA (e.g. SIGReg) scale and compete with SOTA foundation models (DINO)? Here is the answer: yes and with 10x less data. VISReg (slight variation of SIGReg) competes with DINOv2-LVD142M while only training on inet22k. Try it out: https://t.co/vBhrNAmFq6 https://t.co/XERFZEAE8t

📢WorldMesh is accepted to #ECCV2026, and we're releasing the code today! 🎉 Led by @mschneider456: navigable, multi-room 3D scenes from a text prompt, with a mesh scaffold conditioning image diffusion for global consistency + photorealistic detail 👇 https://t.co/8fXCl2flIu https://t.co/Z1HkoO3s37

While we eagerly await Fable 5's return, our agentic WebGPU kernel optimization framework kept running. Opus 4.8 picked up where Fable left off, pushing Liquid AI's new LFM2.5 230M to an unbelievable 1,400 tok/s... running locally in your browser. Don't blink or you'll miss it. https://t.co/27WARZwTcD

1-bit GLM-5.2 GGUF vs. Claude 4.8 Opus vs. GPT-5.5 We gave 3 models the same prompt and compared one-shot outputs. The 1-bit GLM-5.2 GGUF ran locally on a Mac Studio M3 Ultra with 256GB RAM at ~21.6 tok/s. Which output do you like best? GGUF: https://t.co/BMkxswdj5N https://t.co/UoXsCSh4Gn

Easily the biggest unlock for vibe coding 1

At Hugging Face we've been building our own agent that we use via Slack (Moon Bot). Honestly, building your own is quite simple and you'll be happy you did: any model you want (self-hosted if needed), fully customizable to your stack (drop in a skill file and it can use any internal tool, codebase, or DB), your data never leaves your infra, every session auditable in your own bucket. and ofc no lock-in and no waitlist and not overpriced :). read more: https://t.co/7l1w3cib0M

3D scene reconstruction works great until the camera never sees part of the scene. ArtiFixer from NVIDIA Research is an open autoregressive model that fills in the missing geometry that other methods leave blank. #SIGGRAPH2026 paper, code + demo: https://t.co/D9PX2OzbZf https://t.co/AGQicvVKkW

LiteParse is unreasonably good for document parsing ✅ It is the fastest document parsing tool out there - average parse time per page is 3ms ⚡️⚡️ ✅ Now that we support markdown, it tops opendataloader-bench, OlmOCR-bench, and ParseBench in terms of accuracy ✅ It supports 50+ other document formats ✅ It even gives you basic bounding boxes that your coding agent can stitch together Even if you need deeper VLM-enabled parsing (e.g. LlamaParse), there's no reason you shouldn't be using this as a first pass for everything. https://t.co/JNER0mVcB8

+1 more