Breaking

NVIDIA and LangChain Tune Deep Agents Harness for Nemotron 3 Ultra

July 8, 2026 at 12:43 EDT

AI Agents
Foundation Models
Open Source

NVIDIA and LangChain said they have tuned LangChain's long-running agent harness, Deep Agents, for NVIDIA's large model Nemotron 3 Ultra, producing a fully open and customizable agent. The companies say it delivers frontier-class performance comparable to leading closed models at a fraction of the cost.

June 2026 · NVIDIA × LangChain

Frontier-Class Agents, One-Tenth the Cost

NVIDIA and LangChain tuned the Deep Agents harness for the open Nemotron 3 Ultra model — matching leading closed models on accuracy while slashing the price of long-running, multi-step agent workloads.

0.86

Deep Agents aggregate score at just $4.48 per run

5.9×

Inference throughput vs GLM-5.1-754B (8k in / 64k out)

550B

Total params · 55B active MoE, up to 1M-token context

Cost for comparable accuracy

Same task, roughly 10× the price gap

$4.48

Nemotron 3 Ultra
+ Deep Agents

$43.48

Closest competitor
comparable accuracy

Each block ≈ $4.48. Matching accuracy costs almost 10× more on the nearest closed model.

The agentic loop it's built for

Plan

→

Use tools

→

Delegate to sub-agents

→

Recover from errors

Many-turn workflows explode token use — where efficiency cuts task-completion cost by up to 30% on SWE-bench and Terminal-Bench.

Where it shines

Long-running coding, research & internal workflows
Multi-agent coordination close to real-world needs
Cheap iteration on evaluation and tuning
67% Terminal-Bench 2.0 · 82% IFBench

Worth watching

First-hand data still limited post-release
Some code-review tests showed baseline-level results
Not universally best across every task

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.