Breaking

RadixArk Releases Miles: Open-Source Framework for Scaling Reinforcement Learning to Frontier AI Models

September 17, 2026 at 12:02 EDT

Foundation Models
AI Agents
Open Source

RadixArk has released Miles v0.1, an open-source reinforcement learning framework purpose-built for large-scale post-training of frontier language and vision models. The framework integrates SGLang for high-throughput generation, Megatron-LM for distributed training, and Ray for orchestration into a unified stack that eliminates the traditional engineering bottleneck of coordinating rollouts, training, and weight synchronization across thousands of GPUs.

OPEN SOURCE RL FRAMEWORK

Miles Brings Frontier-Scale RL Training to Open Source

Asynchronous rollout and training, optimized weight syncing, and MoE-aware innovations scale reinforcement learning to 744B models across 64 GPUs.

Core Architecture

Rollout (SGLang)
High-throughput token generation with routing capture for MoE models

Training (Megatron-LM)
Distributed training with low-precision (NVFP4, MXFP8) and optimizer offload

Weight Sync (P2P)
Peer-to-peer updates with delta compression (2% for BF16)

Orchestration (Ray)
Distributed task scheduling and fault tolerance

Reference Implementation: GLM-5.2 744B

4.5 min

Training Step Time

64 GPUs Hardware

96% Prefix Cache Hit

Weight Synchronization Speedup: 53.3 sec → 7.2 sec (7.4× faster for trillion-parameter models via P2P transfer)

Supported Model Families

DeepSeek-V4

GLM-5.2

Qwen-3.8

Kimi-K3

Inkling

Nemotron Ultra

Enterprise Context

45.5% of enterprise AI decision makers cite high computational costs as a barrier to RL adoption. 55.4% identify production reliability as a key concern. Miles addresses both through composable architecture and fault tolerance, integrating with existing Megatron-LM stacks rather than requiring system replacement.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.