Is Miles free to use commercially?

Miles is published under the Apache-2.0 license, which permits commercial use, modification and redistribution as long as the license and notices are kept. RadixArk describes the framework as enterprise-facing and says it runs in production at labs and companies, so there is no separate paid edition documented in the repository or on the project site.

What hardware does Miles support?

Miles lists support for NVIDIA GB300, GB200, B300, B200, H200, H100 and A100 accelerators, and for AMD MI300X, MI325, MI350 and MI355X. Supporting both vendors is unusual for reinforcement-learning training code, which normally assumes CUDA only, and it matters for teams whose clusters are AMD-based.

How is Miles related to slime?

Miles began as a fork of slime and, in RadixArk's own words, co-evolves with it rather than replacing it. The technical report frames Miles as the production-hardened path: full-stack rollout engines, several trainer backends, fault tolerance and low-precision training built on that shared foundation.

Which models does Miles support on day one?

Miles ships day-zero configurations for DeepSeek-V4, Kimi-K3, GLM-5.2, Inkling and Nemotron, and also covers Kimi-K2.6 and Qwen3.5. The technical report's case study post-trains a GLM-5.2 744B-A40B model. Dense, mixture-of-experts and multimodal architectures are all covered.

RadixArk · 2026-09-08 · major

Miles v0.1 — RadixArk publishes the technical report for its open RL stack

Miles is an Apache-2.0 reinforcement learning framework for post-training large language and vision models. RadixArk published the v0.1 technical report on September 8, 2026. The repo has 2,685 stars and trains on NVIDIA and AMD accelerators.

GitHub card for the radixark/miles reinforcement learning framework

An Apache-2.0 reinforcement learning framework for post-training large models, now with a full technical report behind it.

Key specs

GitHub stars	2,685

Quick facts

Maker	RadixArk
Version	v0.1
License	Apache-2.0
Training backends	Megatron-LM and PyTorch FSDP2
Rollout engine	SGLang
Hardware	NVIDIA and AMD accelerators
Recipes	GRPO, GSPO, PPO, REINFORCE++, SFT, on-policy distillation

What is it?

Miles v0.1 now has a written technical report: RadixArk posted a 12-author paper on September 8, 2026 describing the system it uses for reinforcement-learning post-training of language and vision models. Miles pairs SGLang for generating rollouts with Megatron-LM for training, and it started life as a fork of the slime project, which the repo says it still co-evolves with. The code is Apache-2.0 on GitHub and had 2,685 stars at the time of this sweep.

How does it work?

Rollout workers and training workers run fully asynchronously and pass tokens to each other directly, so text never gets detokenized and retokenized between the two halves — the repo calls this token-in-token-out. Updated weights move between workers over P2P RDMA, which is what makes trillion-parameter models practical to refresh mid-run. Low-precision paths cover MXFP8, NVFP4, FP8 and INT4 quantization-aware training, and the engine restarts itself when a worker dies.

Why does it matter?

Teams that want to run GRPO, GSPO, PPO, REINFORCE++, supervised fine-tuning or on-policy distillation currently stitch together several projects to do it at scale. RadixArk ships those recipes in one framework with day-zero configs for DeepSeek-V4, Kimi-K3, GLM-5.2, Inkling and Nemotron, plus older Kimi-K2.6 and Qwen3.5. It runs on AMD MI300X-through-MI355X parts as well as NVIDIA, which is rare for RL training code.

Who is it for?

ML infrastructure teams doing RL post-training

Frequently asked questions

Is Miles free to use commercially?: Miles is published under the Apache-2.0 license, which permits commercial use, modification and redistribution as long as the license and notices are kept. RadixArk describes the framework as enterprise-facing and says it runs in production at labs and companies, so there is no separate paid edition documented in the repository or on the project site.
What hardware does Miles support?: Miles lists support for NVIDIA GB300, GB200, B300, B200, H200, H100 and A100 accelerators, and for AMD MI300X, MI325, MI350 and MI355X. Supporting both vendors is unusual for reinforcement-learning training code, which normally assumes CUDA only, and it matters for teams whose clusters are AMD-based.
How is Miles related to slime?: Miles began as a fork of slime and, in RadixArk's own words, co-evolves with it rather than replacing it. The technical report frames Miles as the production-hardened path: full-stack rollout engines, several trainer backends, fault tolerance and low-precision training built on that shared foundation.
Which models does Miles support on day one?: Miles ships day-zero configurations for DeepSeek-V4, Kimi-K3, GLM-5.2, Inkling and Nemotron, and also covers Kimi-K2.6 and Qwen3.5. The technical report's case study post-trains a GLM-5.2 744B-A40B model. Dense, mixture-of-experts and multimodal architectures are all covered.

Try it

docker pull radixark/miles:latest