Nvidia Uses Factory Droid to Train Nemotron
Nvidia used Factory's Droid coding-agent harness to generate about 20,000 rows of post-training data for its 550-billion-parameter Nemotron 3 Ultra model.
Updated on Nov. 21, 2026 in Artificial Intelligence

Nvidia used Factory's Droid coding-agent harness to generate roughly 20,000 rows of post-training data for Nemotron 3 Ultra, an open model with 550 billion parameters released in June. Nemotron 3 Ultra's Terminal-Bench 2.0 score rose from 38.5% to 49% after the post-training work.
Why it matters
The effort shows how a coding-agent harness can surface reproducible failures in tool calling and error recovery, and feed those failures back into model training. The approach points to a method for improving open models on agentic tasks rather than relying only on general-purpose training data.
Nemotron 3 Ultra's Terminal-Bench 2.0 score rose from 38.5% to 49% after training on about 20,000 rows generated through Factory's Droid harness. A preliminary Nemotron 3.5 Super VL checkpoint improved from 52.11% to 60.4% on Terminal-Bench 2.1, though the source does not report results from an independent evaluator.
The players
Nvidia
Nvidia is a American semiconductor and AI company that develops GPUs and the open Nemotron family of language models.
Factory
Factory is a software company that builds Droid, a coding-agent harness used to run and evaluate AI coding agents.
The details
Factory instrumented its Droid harness to capture reproducible failures in instruction following, tool calling and error recovery, and found problems with tool-call and argument formatting in Nemotron. Nvidia then applied PivotRL, a reinforcement learning method introduced in March that trains on critical decisions within existing supervised fine-tuning trajectories without full multi-turn rollouts.
Timeline
March: PivotRL reinforcement learning method introduced.
June: Nvidia released Nemotron 3 Ultra.
Week of November 21: Factory published its case study on the Nvidia Droid deployment.
The Tech Race
The 38.5% to 49% gain on Terminal-Bench 2.0 places Nemotron 3 Ultra in an ongoing contest among open and proprietary models to score higher on agentic coding benchmarks. The result follows a broader push by model developers to use agent harnesses and reinforcement learning to improve tool use and error recovery.
Developers using open models for coding agents may see better tool-call handling and fewer failed recoveries if these training methods are adopted in future model releases. The source does not state release dates or pricing for any updated Nemotron version.
The takeaway
Training on failures captured from real agent runs can improve open models more directly than generic data alone. Readers following AI tools should watch for independent benchmark results before drawing conclusions about model performance.
Further reading
For broader context on model training methods, see Artificial Intelligence coverage on Wisevoter.
Source note: This article includes information reported by The New Stack.



