RadixArk

SHIP AI FOR ALL.

From the core maintainers of SGLang @sgl_project. Building AI infrastructure for inference and training.

[RadixArk @radixark]

We’re proud to support the SGLang Summit as a Platinum Sponsor. We look forward to sharing updates from RadixArk at the summit and connecting with the @sgl_project community.
Join us at Fort Mason in San Francisco, November 12–13!

[SGLang @sgl_project]

Whatever AI infra means to you, we’ve got it covered.
See for yourself: the SGLang Summit agenda is live → https://t.co/7C6TzrkJYV

🗓️ 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?

[RadixArk @radixark]

Love seeing ideas from Miles' Token-In-Token-Out work inspire new research!

Instead of just testing that training sees the exact tokens the model generated, Sasha takes it a step further, formally verifying key properties of a Rust tokenizer in Lean. That's the kind of rigor agentic RL needs.

Fascinating work, @srush_nlp!

[Sasha Rush @srush_nlp]

Lean Verified Tokenizers (https://t.co/sX4zCF4C9y)

Training-inference mismatch kills RL runs, but yet we build messages in Jinja? Prototypes a Lean-verified Rust agent pipeline.

[RadixArk @radixark]

Awesome work from the humans& team!
@humansand's W4A16 NVFP4 work delivers 10% higher throughput and 7% lower TPOT in their SGLang serving benchmark at concurrency 8.

Love seeing SGLang and Miles helping push the boundaries of inference and RL, and seeing the innovations flow back to the open-source community.

Congrats to everyone involved!

[humans& @humansand]

We accelerated one of the most optimized NVFP4 decode kernels while also making it more accurate

Fewer bits typically mean more speed but less accuracy. We did the opposite: more bits, but up to 10% faster generation

Already used in Ampersand’s RL and Boba serving

[RadixArk @radixark]

https://t.co/MhmnTsUTxG

[RadixArk @radixark]

Miles v0.1.3 is out.

This release brings lighter weight updates, a smaller training footprint, and more resilient runs.

Rebuilt P2P transfer cuts the first weight update from 99.5s to 3.9s and pinned memory per sender from 28 GiB to 2 GiB in our benchmark.

Fully async RL can now share GPUs between training and rollout, even on a single GPU. The new fused log-prob backend cuts peak GPU memory by about a third in our long-response test.

Failures stay contained. With fault tolerance, a failed cell no longer blocks weight updates to healthy cells. Fully async runs can also resume where they left off.

LoRA now runs in Megatron raw mode across model families using a unified implementation.

Full release notes 👇

[RadixArk @radixark]

We trained the DFlash2 draft for MiniMax M3.1 Flash, ready to use in SGLang from Day 0 🐰
M3.1 Flash also has Day-0 RL training support in Miles!

Congrats @MiniMax_AI! Great working with @sgl_project, @NVIDIAAI and @AMD on this launch.

[SGLang @sgl_project]

Space Bunny is here! 🐰 MiniMax M3.1 Flash hits >1K tok/s, bs1 decode with SGLang, production-ready from Day 0. This integration is a joint effort by @MiniMax_AI, @NVIDIAAI, @AMD, and @radixark, and you can also run RL with Miles!

-Broad hardware support: NVIDIA Blackwell and Hopper (GB300/B300/B200/H200), plus AMD Instinct MI350X and MI355X.
-Agentic serving at scale: ~100K total tok/s per GPU on the AgentX coding-agent trace with 96 concurrent sessions.
-Speculative decoding: Supports MiniMax's DSpark draft and a DFlash2 draft trained by @radixark; supports HiCache and multimodal input.
-Ultra-low-latency inference: MegaKernels deliver 2.15× faster single-request decode.

[RadixArk @radixark]

Heterogeneous RL with Miles: generate rollouts on AWS Trainium, train wherever your policy lives. Pick the hardware for each phase, and scale rollout capacity on its own.

This is an early preview, with more to share at SGLang Summit next month 👇
https://t.co/HlVRywiBNV

[SGLang @sgl_project]

Today, in close collaboration with @awscloud and alongside the TorchNeuron launch, we’re announcing the beta of the SGLang AWS Trainium plugin, built on TorchNeuron and running on Trainium2 and Trainium3.

What’s in the beta:
- Inference serving and RL rollouts
- Support for Qwen3, DeepSeek, and GLM model families, with more frontier models on the way
- Early preview of heterogeneous RL, with rollouts on Trainium and training via Miles (@radixark)

Stay tuned for more at the SGLang Summit next month!

[RadixArk @radixark]

SGLang-TorchTPU gives people options to choose the hardware that best fits their workloads and economics. We're excited to see many partners already running it in production!

Huge thanks to everyone at Google, Meta, our @sgl_project team, and community who made this happen!

We're committed to helping make TPU a first-class citizen in the SGLang ecosystem and will keep building toward a more flexible, heterogeneous future for AI infrastructure.

[SGLang @sgl_project]

Today we launch SGLang-TorchTPU, built with @Google TPU team and Meta @PyTorch team.

We’re excited to see it powering production inference at Meta!

SGLang-TorchTPU is a TPU plugin for SGLang that brings:

- Top open models on TPU: Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash, MiniMax H3
- SGLang’s production features for large agentic workloads
- Custom MegaKernel: Muse Spark Open hits 770 tok/s decode at bs=1 on InferenceX 8K/1K

SGLang-TorchTPU is available now: https://t.co/ix02YggKGG

[Reflection @reflection_ai]

Model customization recipes are live on our website and in the cookbook repo.

Beam is now supported across RL and SFT stacks including @axolotl_ai, NeMo RL, NeMo AutoModel, Megatron Bridge (@NVIDIAAI), Miles (@radixark), SkyRL (@NovaSkyAI), and TRL (@huggingface).

Get cooking 🧑‍🍳
https://t.co/gRHcqEsa0s

[RadixArk @radixark]

Beam × Miles 🤝
We've published RL post-training recipes for @reflection_ai's Beam, covering full-parameter GRPO and LoRA training with SGLang-powered rollouts.

Thanks @chetantekur and the Reflection team for the collaboration!

Check out the cookbook: https://t.co/at8CIKBj8L

[Chetan Tekur @chetantekur]

We have been working on making Beam easier to customize at @reflection_ai. Seeing the recipes and integrations go live feels pretty amazing.

We're putting the code, training configs, and examples out there so people can start building on Beam. Here's what's in the cookbook:

Data generation: Use Beam to create training data, including software-engineering trajectories that pass tests and reasoning examples for tool calling.
https://t.co/nHFVuJTIgx

Fine-tuning: Adapt Beam to your data and tasks with LoRA, QLoRA, or full fine-tuning.
https://t.co/kyhY5O4pB5

Distillation: Use Beam to teach a smaller model. We've included recipes for coding trajectories and tool-calling reasoning.
Coding: https://t.co/cpWNS1ItNh
Tool calling: https://t.co/GDxnHcecpD

Reinforcement learning: Post-train Beam against verifiable rewards.
https://t.co/ud9CjW29er

Huge thanks to our partners and the teams behind these integrations:

• NeMo AutoModel, @NVIDIAAIDev: SFT with LoRA and full fine-tuning.
https://t.co/DFHj1Njy2D

• Megatron Bridge, @NVIDIAAIDev: Large-scale full fine-tuning.
https://t.co/ugdY0z5dG4

• Axolotl, @axolotl_ai: LoRA, QLoRA, and full fine-tuning.
https://t.co/5MIEEUi4Wb

• HF TRL, @huggingface: Training smaller models on Beam-generated data.
https://t.co/18CR3KIo6g

• Miles, @radixark: RL post-training.
https://t.co/fFPHGLr3Bu

• NeMo RL, @NVIDIAAIDev: RL post-training on NVIDIA hardware.
https://t.co/H42WCcCrVd

• SkyRL, @NovaSkyAI: LoRA and full-parameter RL.
https://t.co/usmEpQlVH3

Each recipe has its own instructions, dependencies, sample inputs, and smoke test. I'm excited to see people run these experiments, adapt Beam to their own work, and contribute recipes we haven't thought of.

Huge thanks to @Nazeri2010, @bhutanisanyam1, @kathleen_mhatre, @FGbelidji, @joespeez, and to Thibaut, Matthias, Travis, Kai, and Yoni for making this happen.

Explore the cookbook: https://t.co/rbGXC4mzFY

Access the model: https://t.co/hKPmuJpQcZ
Documentation: https://t.co/LEfcF98AYV
Provide feedback on Discord: https://t.co/98piDyBK3g

[PyTorch @PyTorch]

SGLang-Diffusion is a high-performance serving framework for diffusion models, designed for both large-scale offline generation and latency-sensitive real-time inference.

At PyTorch Conference North America 2026, Yihao Wang and Kevin Mi of @RadixArk will present the design of @sgl_project-Diffusion and the PyTorch-based inference optimizations behind it, including efficient request scheduling, memory management, batching strategies, and execution-path optimizations for diffusion pipelines.

Join us in San Jose, October 20-21: https://t.co/jBApW8ocHQ #PyTorchCon

[RadixArk @radixark]

We’re at #PyTorchCon 2026 in San Jose, Oct 20–21 👋

Catch the talks from the RadixArk team:
• SGLang-Diffusion (Yihao Wang, Kevin Mi)
• Native PyTorch RL on AWS Trainium (@alxnails, with Yahav Biran from Annapurna Labs)
• Miles for agentic RL (@yueming02)
Session times and speakers below. Come say hi! @PyTorch @sgl_project

[RadixArk @radixark]

@reflection_ai Miles cookbook: https://t.co/at8CIKBj8L

[RadixArk @radixark]

Our team worked with @reflection_ai to get Beam running on SGLang, and it now serves their production API. Beam is also supported in Miles for RL training. Congrats to the Reflection team on the launch!

[SGLang @sgl_project]

SGLang powers Beam’s production API, with day-0 support now available to everyone!

Built for software engineering and long-horizon agents, Beam delivers frontier-level inference efficiency with more intelligence per token. Serve it on the same production-proven stack: https://t.co/KlXMFtR7Oa

Read more: https://t.co/73ACF53DQr

[RadixArk @radixark]

@AIatMeta @sgl_project @NVIDIAAI @AIatAMD @Google Miles cookbook: https://t.co/jCXTf7zUV1

[RadixArk @radixark]

Great working with @AIatMeta to bring production-ready support for Muse Spark Open to @sgl_project and Miles.
Meta’s custom, auto-tunable Helion kernels deliver 1.03–3.95× kernel-level speedups across @NVIDIAAI and @AIatAMD GPUs. Together, we also added day-0 support on @Google TPU via SGLang-TorchTPU for the first time.
Proud to work with Meta on this launch and bring the work we do at RadixArk to the SGLang and Miles communities.

[SGLang @sgl_project]

Muse Spark Open is built to plan, act, and use computer across agentic tasks for life and work.

You can now build a powerful personal AI agent with an open stack, and SGLang to make it:

- Ultra-responsive with NVFP4 MegaMoE and sequence parallelism
- Faster for agentic and coding tasks with our DFlash2 draft model
- Customizable for specialized use cases via fine-tuning with Miles

We’re incredibly proud to have made this possible with @AIatMeta, @NVIDIAAI, @AIatAMD, @intel, and @radixark!

[RadixArk @radixark]

@StepFun_ai Miles cookbook: https://t.co/WUsUeeaJYM

[RadixArk @radixark]

Congratulations to @StepFun_ai on Step 5 Preview going open weights! RadixArk is proud to help power StepFun's first-party production API with SGLang at scale.

Production inference means great token economics under strict latency requirements. SGLang and its cache-aware routing stack now bring that to Step 5 Preview in proven production traffic.

Building with Step 5 Preview? Deploy with SGLang today!

[StepFun @StepFun_ai]

Step 5 Preview is now open weights.

BF16 and FP8 weights are live on Hugging Face and ModelScope, together with our technical blog.

• Sparse MoE: 600B total, 27B active per token
• 1M-token context window
• Native text and vision input
• Built for long-horizon agentic work: coding, software engineering, professional knowledge work and finance

Frontier intelligence, open weights. Build what's next.

Hugging Face: https://t.co/nMfayf37ja
ModelScope: https://t.co/mkB3qIoMQA
GitHub: https://t.co/siRkQqaruX

[RadixArk @radixark]

https://t.co/GUAiOv4Yos

[RadixArk @radixark]

Miles-Diffusion now supports RL training for FLUX 3 Image Klein Base from @bfl_ai.

The whole pipeline runs on a single node. Both LoRA and full-parameter RL have been validated end to end.

Cookbook in the comments 🔗

[Black Forest Labs @bfl_ai]

FLUX 3 Image Klein is here.

Open Weights with the quality of a large frontier model. Customizable and fast.

5.6B parameters and runs on a single consumer GPU.

Native 4K, bounding boxes, and pixel-perfect editing.

[RadixArk @radixark]

Cookbook: https://t.co/0oVH80JEw2