Breaking

NVIDIA and LangChain release agent blueprint tuned for Nemotron 3 Ultra

July 8, 2026 at 12:47 EDT

AI Agents
Foundation Models
Open Source

NVIDIA and LangChain on July 8, 2026 released the "NemoClaw Deep Agents Blueprint," which optimizes LangChain's Deep Agents harness for the open model Nemotron 3 Ultra. By tuning the harness rather than the model itself, they claim to deliver performance on par with leading closed models at a fraction of the cost.

July 8, 2026 · NVIDIA × LangChain

Frontier-Class Agents at a Tenth of the Cost

The NemoClaw Deep Agents Blueprint tunes the agent harness—not the model weights—for the open Nemotron 3 Ultra, matching leading closed models while slashing the price of running long tasks.

0.86

Aggregate agent-eval score

~10×

Lower cost vs next-best model

Token context window

Cost for comparable performance

Same 0.86 score on the agent evaluation suite — price per run.

$4.48

NemoClaw
Blueprint

$43.48

Next-best
model

Same result. Roughly one-tenth the bill.

Nemotron 3 Ultra — agentic benchmarks

95%

RULER @1M

91%

PinchBench

82%

IFBench

67%

Terminal-Bench 2.0

550B total / 55B active params

MoE · hybrid Mamba-Transformer · NVFP4

5–5.9× higher throughput

vs competing open & closed peers

What developers like

Frontier-class agents at far lower cost
Fully open stack they can own & modify
~300 tokens/sec; 56–62% pass on code-review

Caveats in production

Elevated retry rates in some evals
Gap can remain vs top closed models
Harness & runtime tuning proves decisive

Continue reading

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