Miles v0.1 launches as open-source RL framework for large language and multimodal models

RadixArk announced the release of Miles version 0.1, an open-source reinforcement learning (RL) framework intended for training large language models and multimodal models. The tool addresses a well-known problem for RL teams: starting a run is easy, but debugging when something goes wrong is a different story. Miles is built to ensure runs are correct, to use hardware efficiently, and to keep training stable at large scale.

Development lasted nine months and involved 72 contributors who submitted 1,326 commits and 85 end-to-end CI tests on GPU. The framework was field-tested on advanced open-source models including Kimi K3, DeepSeek V4, Qwen 3.8, GLM 5.2, Inkling and MiniMax H3. Organizations already running production and research workloads on Miles include Humansand, Periodic Labs, Modal, DecagonAI, Eigent.AI, Nebius AI and IBM, using hardware from NVIDIA and AMD.

Periodic Labs builds training of trillion-parameter models on thousands of GPUs with especially long-horizon scientific RL rollouts. According to the company, the integration delivered a three-fold improvement in rollout throughput, a ten-fold improvement in weight synchronization and a thirty-fold improvement in weight conversion. Humansand uses the framework as core infrastructure for asynchronous multi-agent long-horizon RL research and adapted it to its own orchestration with low-precision RL recipes MXFP8 and NVFP4. Eigent.AI leverages trajectory management and the rollout infrastructure to train agents across different models, tasks and environments. DecagonAI runs post-training of AI agents through Miles for instruction tracking and tool activation, aiming for low-latency, high-accuracy customer experiences.

Miles ships with connectors in the rollout layer for a range of training environments: Harbor, HUD, NVIDIA’s NeMo Gym, Hugging Face’s OpenEnv and Prime Intellect’s Verifiers. Sandboxes run on AgentENV. This means a team can switch training environments without rewriting collection and synchronization infrastructure, a common pain point in projects that combine multiple benchmarks or simulators.

The launch signals a maturation of RL infrastructure for large models: less glue code that each team writes themselves, more shared components that have passed CI on real hardware. The version is still labeled 0.1, indicating an unstable API and expected breaking changes, but the reported efforts and metrics suggest the direction is sound. For teams weighing whether to build their own solution or adopt one, Miles offers a tested starting point at a scale many labs cannot achieve on their own.