FULL STORY

NVIDIA's Nemotron Ultra: From Tuning to Open-Source Viral Hit

NVIDIA partnered with LangChain to tune the open-source Nemotron Ultra model, which subsequently went viral on the Ollama platform, highlighting NVIDIA's strong commitment to the open-source community.

2026-07-08 ~ 2026-07-14 · 2 episodes · 9 posts

Episode 1 · NVIDIA and LangChain Tune Deep Agents: Nemotron 3 Ultra Nears Opus (2026-07-08, 6 posts)

NVIDIA and LangChain have partnered to tune the open-source Deep Agents harness. This collaboration highlights a deep integration in agent frameworks and engineering stacks, demonstrating the high cost-efficiency potential of open-source models through concrete benchmark data.

Key Details and Performance Comparison

According to LangChain's internal deep-agents benchmark, NVIDIA's Nemotron 3 Ultra achieved a composite score of 0.86 (86%). In comparison, Claude Opus 4.8 scored 0.87, meaning the two models are separated by just 1 percentage point in performance. However, on the cost side, Nemotron 3 Ultra's inference cost was only $4.48. Several users, such as @BraceSproul, emphasized that with Deep Agents support, the model's usage cost has been reduced to a fraction of the original, enabling faster AI workflows and stronger decision-making.

Reactions and Industry Impact

In his repost, @hwchase17 pointed out that AI progress is not dictated solely by the release cadence of cutting-edge closed-source models; open-source models and their accompanying systems are improving rapidly as well. NVIDIA's official account also confirmed that this tuning work is now public. This collaboration and the subsequent evaluation results visually demonstrate that open-source agent stacks can now approach the performance of top-tier closed-source models with significantly better cost-efficiency.

Episode 2 · NVIDIA's Open-Source Nemotron Ultra Surges in Popularity (2026-07-14, 3 posts)

NVIDIA's newly open-sourced Nemotron Ultra model is experiencing rapid growth on Ollama. The release includes comprehensive open-source assets like model weights and training code, aiming to empower developers with complex, long-running agentic tasks.