AI News

Trending story

RadixArk and NVIDIA Run Miles Agent RL Training End-to-End on Vera CPU

1 reports1 story updatesUpdated Nov 13 08:00

Understand this story

AI overview

RadixArk has partnered with NVIDIA to run all sandbox executions of the open-source RL post-training framework Miles on NVIDIA Vera CPUs, which it says is the first publicly reported result of an end-to-end agent RL training sandbox deployment on Vera CPU servers. In a comparison at full load of 176 threads, each episode on Turin was on average 12% slower and took 53 seconds longer, with agent commands taking 60% more time. The team used Miles to train GLM-4.7-Flash on Terminal-Bench and SWE-bench tasks, for a total of 68 steps and over 2,400 episodes, with the pass rate on gRPC service tasks rising from 40% in the first half of training to 82% in the second half.

AI-generated from public reports

Latest development · Report

RadixArk Miles Runs Agentic RL Training on NVIDIA Vera CPUs: Faster Sandboxes, Faster Training

Developments

1 confirmed developments

  1. RadixArk and NVIDIA run Miles agent RL training end-to-end on Vera CPU

    Source:LMSYS Blog

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 13

  1. LMSYS BlogFirst-party

    RadixArk Miles Runs Agentic RL Training on NVIDIA Vera CPUs: Faster Sandboxes, Faster Training

    RadixArk partnered with NVIDIA to run all sandboxes of the open-source RL post-training framework Miles on NVIDIA Vera CPUs, which it calls the first publicly reported result of an end-to-end agentic RL training sandbox deployed on Vera CPU servers. In a comparison at full load of 176 threads, each episode on Turin was on average 12% slower and took 53 seconds longer, with agent commands taking 60% more time. The team used Miles to train GLM-4.7-Flash on Terminal-Bench and SWE-bench tasks, for a total of 68 steps and over 2,400 episodes, and the pass rate on gRPC service tasks rose from 40% in the first half of training to 82% in the second half.

Related stories

  1. Related storiesCohere discloses that the Qwen3.8 model attempted to exfiltrate files in a sandbox
  2. Related storiesLMSYS uses Miles to train SWE agents on MI355X
  3. Related storiesNvidia releases Nemotron 3.5 Super VL and Bolt models
  4. Related storiesEnglish translation is being prepared.