v0.1.0

Miles v0.1.0 is the first versioned release of Miles — a full-stack, production-ready system for frontier post-training, built on slime's foundation and verified, clean, and customizable everywhere.

Read the release blog: https://lmsys.org/blog/2026-08-18-miles-v0-1

Highlights

• Fully asynchronous RL — eliminates the synchronization bottleneck between rollout and training; the GLM-5.2 744B reference run sustains ~4.5-minute training steps on 64 GB300 GPUs with a 96% prefix-cache hit rate.
• Token-In-Token-Out (TITO) session server — preserves exact tokenization across multi-turn agentic sessions.
• Rollout Routing Replay (R3) — tames MoE sensitivity to numerical differences between rollout and training.
• Low-precision training — NVFP4, MXFP4, MXFP8, FP8, and INT4 QAT recipes.
• P2P (RDMA) weight transfer — cuts weight-update time from 53.3s to 7.2s on 1T-parameter models; disk-delta updates shrink the payload from 62.4 GB to 0.69–0.83 GB.
• NVMe optimizer-state streaming for memory efficiency on large models.
• LoRA RL for dense and MoE models — GLM-5.2 744B LoRA trains with 212 MB of trainable parameters (0.014% of the BF16 base).
• On-Policy Distillation (OPD) and Zero-KL alignment between training and rollout.
• Environment integrations — Harbor, HUD, NeMo Gym, OpenEnv, Prime Intellect Verifiers.

Beyond LLMs, RL for flow models (SD3.5, Qwen-Image, LTX-2.3, Cosmos3, Minimax-H3) is supported in miles_diffusion.

Models

Kimi-K3, DeepSeek-V4, Inkling, Qwen3.8, NVIDIA Nemotron 3 Ultra, DeepSeek-V3.2, Kimi-K2.6, GLM-5.2, Gemma-4, GPT-OSS, and more.

Hardware

NVIDIA H100, H200, B200, B300, GB200, GB300; and AMD MI300X, MI355X.

Docker

docker pull radixark/miles:v0.1.0        # CUDA 13
docker pull radixark/miles:v0.1.0-cu12   # CUDA 12

All contributors

We would like to thank all community contributors of Miles and SGLang for their invaluable contributions and continued support.

All Miles contributors (in alphabetical order):

@0x404 @AgainstEntropy @Arist12 @ArtificialZeng @Aurelius84 @austin362667 @bcol23 @BHZ-BER @ByronHsu @casper-hansen @Chen-GX @ChenyangWangjohn @dougyster @FrankLeeeee @fy1214 @fzyzcjy @gali-leilei @Gao016 @GeeeekExplorer @GeLee-Q @GGGGGGXY @gongyisheng @gramesh-amd @GuanxingLu @guapisolo @Guido1Alessandro1Trevisan @Hallboo @harvenstar @hebiao064 @Hecate0821 @HJSang @hyleepp @indianspeedster @Irvingwangjr @JD-ETH @JensenFire @JessicaJiang-123 @jhinpan @jthomy @JustinTong0323 @kaixih @kakisong @lancerts @lawrence-harmonic @layahaasini @lilei199908 @lizamd @luppx @Lyken17 @maocheng23 @mathewjhan @Maybewuss @Michaelsqj @miles-code-angel @momo609 @mouad-hpc @MrAta @MrWhitezz @nanjiangwill @nblintao @NJX-njx @none0663 @nono-Sang @oraluben @PopSoda2002 @ppraneth @QQontheMoon @Ratish1 @rbao2018 @richardodliu @Risc-lt @Rockdu @rucnyz @rysaya @sam571128 @SanftMonster @Shi-Dong @souhil25 @sreerohi @SwordFaith @techkang @TSunny007 @UbeCc @unseenmars @VincentXWD @waitong94 @willccbb @Williamren97 @WindowsXp-Beta @WWWjiahui @XihWang @Xinyu-Kang @XinyuJiangCMU @xiuhu17 @yefei12 @Ying1123 @yinpeisu @yitianlian @ylong18 @ys-2020 @YuchenFan48 @yueming-yuan @yushengsu-thu @zhaochenyang20 @Zhichenzzz @Zhuohao-Li @zhuzilin @zianglih @zijiexia @zkx06111 @zlH518 @zRzRzRzRzRzRzR @zxpdemonio @zyzshishui