AgileRL

Build, train and ship agents. Powered by reinforcement learning.

[AgileRL @AgileRL_Inc]

We are proud to announce a collaboration between AgileRL and @NVIDIAAI to support the post-training of Nemotron models. NVIDIA's Nemotron open-source model family is now available for fine-tuning on Arena, our platform for creating AI agents specialized at any task.

NVIDIA gave us early access to their latest 30B3A parameter MoE model, Nemotron 3.5 Lightning. On Arena, we trained it to outperform Claude Sonnet 5 on two different tasks: a long-horizon reasoning challenge and a real customer support workload. Nemotron proved exceptionally easy to post-train: a single six-hour run on four H100 GPUs was enough to master the reasoning task, at context lengths beyond 50,000 tokens. The full results are in the announcement, linked below.

With this collaboration, Arena customers get access to NVIDIA Nemotron models as soon as they are released, with the training and evaluation setup already built around them. Post-train Nemotron on your own data, environment and edge cases, to build an agent that masters your task.

Across industries including banking, insurance, aerospace and government, production traffic is moving off frontier model APIs and onto infrastructure these companies control. Businesses want agents with genuine expertise in their specific task, trained on their own data. Security, control and data sovereignty demand that model weights remain on their own infrastructure. And they want the fixed cost of hardware they own, rather than per-token pricing that compounds with every request.

This collaboration provides that path. Begin with a dataset, an RL environment, or simply a description of the job the agent has to do, and our team will build the rest with you. Arena handles the training and deploys the finished agent in one click, with the weights yours to keep.

[Fusion Fund @FusionFundVC]

Congrats @AgileRL_Inc & @kumar_param927! With @nvidia Nemotron 3.5 Super VL, AgileRL hit 90.2% success on browser tasks vs 63.2% for @claudeai Opus 5.5. #ReinforcementLearning can tune #OpenModels to surpass frontier models at real-world computer use.

🔗: https://t.co/AVLxDInVdr

[NVIDIA AI @NVIDIAAI]

Love seeing our partners post-training Nemotron 3.5 Lightning for their own domains, tools, and workflows, as well as those offering post-training support ⚡

Check out below to see what they’re building 🧵

[AgileRL @AgileRL_Inc]

@NVIDIAAI Excited to be an NVIDIA ecosystem partner. Post-train the Nemotron model on our platform, Arena.

Learn more about the partnership here: https://t.co/knMH1taLmx

[NVIDIA AI @NVIDIAAI]

Introducing NVIDIA Nemotron 3.5 Lightning⚡

An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster.

It delivers up to 4x the output speed of similar-sized models.

[AgileRL @AgileRL_Inc]

Blog post: https://t.co/knMH1taLmx

NVIDIA announcement: https://t.co/kkMBMOMMJl

[AgileRL @AgileRL_Inc]

We are proud to announce a collaboration between AgileRL and @NVIDIAAI to support the post-training of Nemotron models. NVIDIA's Nemotron open-source model family is now available for fine-tuning on Arena, our platform for creating AI agents specialized at any task.

NVIDIA gave us early access to their latest 30B3A parameter MoE model, Nemotron 3.5 Lightning. On Arena, we trained it to outperform Claude Sonnet 5 on two different tasks: a long-horizon reasoning challenge and a real customer support workload. Nemotron proved exceptionally easy to post-train: a single six-hour run on four H100 GPUs was enough to master the reasoning task, at context lengths beyond 50,000 tokens. The full results are in the announcement, linked below.

With this collaboration, Arena customers get access to NVIDIA Nemotron models as soon as they are released, with the training and evaluation setup already built around them. Post-train Nemotron on your own data, environment and edge cases, to build an agent that masters your task.

Across industries including banking, insurance, aerospace and government, production traffic is moving off frontier model APIs and onto infrastructure these companies control. Businesses want agents with genuine expertise in their specific task, trained on their own data. Security, control and data sovereignty demand that model weights remain on their own infrastructure. And they want the fixed cost of hardware they own, rather than per-token pricing that compounds with every request.

This collaboration provides that path. Begin with a dataset, an RL environment, or simply a description of the job the agent has to do, and our team will build the rest with you. Arena handles the training and deploys the finished agent in one click, with the weights yours to keep.

[AgileRL @AgileRL_Inc]

Ships with AgileRL v2.8.0: LocalTrainer and ArenaTrainer load the same manifest, so moving from your laptop to the cluster is a one-line change.

Full walkthrough: https://t.co/zMfmIZ8tHH

Get started: https://t.co/kk22EV4TPl

[AgileRL @AgileRL_Inc]

The Arena Client is out: a Python SDK and CLI for Arena, our managed RLOps platform.

Define and steer RL experiments from your terminal. Training runs on cloud infrastructure built for RL, with population-based training and evolutionary HPO.

pip install agilerl-arena

[AgileRL @AgileRL_Inc]

Two details that matter: manifests and reward functions are validated before any GPU time is spent, and every checkpoint is kept, because reward curves are not monotonic and the best model is rarely the last one.

One command deploys the best checkpoint to a managed endpoint.

[AgileRL @AgileRL_Inc]

CISPO needs the throughput the async architecture delivers, and async is only safe because CISPO tolerates stale data. Neither works alone.

Full write-up: https://t.co/QVOlbNyKjk

Open source: https://t.co/JkimNFlDYI

[AgileRL @AgileRL_Inc]

RL training for LLM agents is expensive because rollout and training take turns sitting idle.

We rebuilt the loop. AsyncRL + CISPO in Arena: 7x cheaper than ART, 5.3x the token throughput of TRL, on half the GPU memory. Both baselines failed to learn the task at all.

How:

[AgileRL @AgileRL_Inc]

Cheap parallel agents mean hyperparameter search moves inside the run. A population of 4 reached in 5 hours the score a single agent took 16 to hit, and cost $472 total.

One failed ART run cost $811.

[AgileRL @AgileRL_Inc]

Hyperparameter tuning is the slowest part of RL. Hundreds of sequential training runs, weeks of waiting, and the "best" config is stale the moment the environment shifts.

AgileRL evolves hyperparameters during training, in a single run. Up to 10x faster than the SOTA.

Read more:
https://t.co/cEI6tSPSMD