← News

Agents & MCPSources: arXiv · 19 Nov 2026

MOLT framework trains Hugging Face models for agentic reinforcement learning at scale

The open-source research framework loads Hugging Face checkpoints directly into a PyTorch trainer, scaling from 4B to 753B parameters without bridge code.

In plain words

MOLT is an open-source tool that helps researchers train AI agents with reinforcement learning, a method where agents improve through rewards. It works directly with Hugging Face model files, so researchers can modify and run the code without extra conversion steps, even for very large models.

What was announced

The authors say existing open frameworks force a trade-off: easily modified ones do not scale to the largest models, while scalable ones depend on Megatron and require bridge code for each model family. MOLT instead loads a Hugging Face checkpoint into a PyTorch FSDP2 trainer with tensor, expert, and context parallelism, so a new model family needs no bridge. Its full reinforcement learning path is roughly 9.2K lines of Python with one training backend and one rollout engine, which the authors say makes it readable and editable by researchers or AI coding agents.

Reported results include a 30B mixture-of-experts tool agent trained stably for 1,464 asynchronous updates, a visual computer-use agent improving on OSWorld, asynchronous training running 2.40 times faster than synchronous training, and pipeline throughput reaching 92% of a Megatron-based stack. The trainer, NeMo AutoModel, is said to handle models from one to 2.8 trillion parameters. Existing OpenAI- and Anthropic-SDK agents can be trained by redirecting their client, with training fully asynchronous and trajectory data held in a distributed object store.

What it does not show

  • All claims come from the paper's own abstract; no independent testing or confirmation is reflected here.

For security and IT teams

Scope
The source is the paper's abstract only, so implementation details beyond the summary are not covered.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.