How much can you improve a small model in 18 days on 144 GPUs?
Meet DaisyRose-27B, RadixArk's first open-weight model! It is trained with RL on Qwen3.8-27B, with 1,200 curated long-horizon agent tasks for 18 days on 144 H200s. Powered by Miles and our own training backend, we stretched agent sessions up to 1M tokens and raised the Terminal-Bench 4.1 score from 5.6% to 12.0%.
See how we pull it off ⬇️
All these dudes crying "whys sarvam hosting Chinese models" need to understand it's a good way to utilize capacity, serving models people will actually use.
Some news: I'm starting a company in the AI for Science space.
We are training foundation models and improving them for multi-agent scientific discovery.
If you're looking to join an early-stage startup as a Research Scientist or Research Engineer, I'd love to chat. My DMs are open!
2,000+ people. Two days. The first SGLang Summit.
Together, we announced SGLang v1.0, released Miles v0.2, and launched the SGLang Consortium with nine founding members: @nvidia, @AMD, @intel, @Google, @Meta, @thinkymachines, @Cloudflare, @modal, and @radixark.
SGLang is now used by 332 organizations. At Meta, it powers production inference for products serving billions of users. Teams from Modal, Cloudflare, LinkedIn, and others shared their production lessons.
Across 90+ technical talks and six hands-on workshops, the community explored what comes next for inference, reinforcement learning, and AI agents.
To everyone building, deploying, and contributing: thank you for making this possible.
Explore the announcements, engineering lessons, and moments from SGLang Summit 2026 in the link below.
More than ever open source should win, and to get there, we also need to make open models easier to use for all!
Today, we're joining forces with @OpenRouter: Hugging Face Inference Providers, our serverless inference API for open models, is now powered by OpenRouter. Every request goes to the best provider, same endpoint, same HF token.
This is the first part of our partnership with OpenRouter, more to come soon!
hf.co/blog/openrouter
This Thursday and Friday, we're hosting the first SGLang Summit, bringing together 2,000+ people from across the AI ecosystem.
When I joined as @sgl_project 's first PM, we were still a small team figuring things out as we went. We hadn't hosted a proper meetup, barely knew anyone in the ecosystem, and spent countless hours in war rooms bringing up support for new models. I still remember someone from a leading hardware company telling me at our first meetup that SGLang accounted for less than 10% of the inference requests they were seeing, mostly DeepSeek workloads.
A lot has changed since then. Over the past year, I've worked with hundreds of teams across model labs, hardware vendors, cloud providers, and AI companies, turning conversations into partnerships, product improvements, and production adoption. Today, SGLang provides day-0 support for releases from all major model labs, with a growing number of companies relying on it for production inference across diverse models and workloads.
This Summit brings so much of that journey together. Many of the partners and production users we've worked with will be sharing their experiences on stage. Seeing the relationships we've built and the ecosystem we've grown come together in one place is incredibly rewarding.
I'm also especially proud of the team we've built along the way. What started with a handful of us wearing many hats has grown into an incredible team of nearly 20 people across Product, Marketing, Growth, Partnerships, and BD. Together, we've driven product and ecosystem growth while bringing this Summit to life. It has been a massive undertaking, and the ownership, dedication, and execution this team has shown have been extraordinary.
From our very first meetup to SGLang Summit, it's been quite a journey. And we're just getting started.
See you at SGLang Summit!
📖 The Ultimate MoE Training Handbook 📖
We spent 1/2M GPUh to optimize MoE training and wrote the most comprehensive blog about it. With 30k words and 50+ interactive figures it covers everything from profiling to large cluster training.
huggingface.co/spaces/huggin…
Kimi K3.1 Pro replaces dense MLA with sparse attention (DSA). It's much more efficient and easier to serve.
SGLang delivers on Day 0:
· Up to 33.7K tok/s per GPU (AgentX, GB300)
· 450 tok/s single user with DFlash2 + MegaKernel
· NVIDIA Vera Rubin / Blackwell / Hopper, AMD, TPU, Trainium, PPU
Our picks from this week's model releases 🦙
→ LFM3 by Liquid AI: small models for text, image and speech, with d1 decision models
→ Mistral Large 4: large, 1M context frontier
→ Flint by Recursive: mid-sized, built for coding agents
Get them at llama.app/models
🥁 I'll be speaking at the SGLang Summit, 2026!
SGLang (@sgl_project) is one of the most critical pieces of infrastructure in AI. Proud to represent Hugging Face at the Summit!
Congrats to @Baidu_Inc on Unlimited OCR 2! An entire book parsed in a single request, with decode latency and memory staying flat across hundreds of pages 👀
Proud to be the recommended serving path. Serving instructions are in the model card: huggingface.co/baidu/Unlimit…
Cookbook coming
Unlimited OCR 2 is here, and it's open source. 📖 ✨
Longer documents can mean a lot more attention computation. Our new SR-SWA mechanism keeps that in check by condensing text-visual attention interactions into a fixed-size set of working references.
Here's what it delivers:
- Document parsing at virtually any length, within a finite context window
- Nearly constant per-step decoding latency and memory footprint as documents grow longer
- State-of-the-art performance on MPDocBench and MarathonDocBench, covering documents from dozens to hundreds of pages
Check out the model here ↓
The prompt isn't asking the LLM to curl, it's asking the LLM to find a way to monitor if other things curl stuff. So if the LLM then starts with a test probe file and sends to a server that was never mentioned in the prompt or reasoning trace that is behaviour that's out of the ordinary. Again, not asserting malicious intent, we're flagging that it went out of bounds of the prompt in a potentially harmful way.
DeepInfer set the bar. SGLang pushed it higher! 🚀
SGLang now matches DeepInfer performance on BOTH prefill and decode:
• Prefill: Up to 1.04× throughput on 4× GB300 and 4× B300
• Decode: Up to 7.4% throughput on 64× GB300
We also provide an end-to-end reproduction guide and a full serving stack to run it in production.
Whatever AI infra means to you, we’ve got it covered.
See for yourself: the SGLang Summit agenda is live → sglang.io/summit
🗓️ Nov 12–13 · 📍 Fort Mason, SF
🎟️ Early bird ends Oct 20
Two days of open models, AI hardware, recursive self-improvement, agents, robotics, and the systems behind them.
Featuring @LipBuTan1 (Intel), @lilianweng (OpenAI), @soumithchintala (Thinking Machines), @AravSrinivas (Perplexity) and researchers and builders from across the AI ecosystem.
Plus hands-on workshops on model training and serving decision models, hardware Expo, and plenty of community events.
What does AI infra mean to you?