NVIDIA Nemotron 3.5 Lightning Matches Larger Models on Terminal-Bench Despite Compact Size

Nemotron 3.5 Lightning is a 31.6B total / 3.6B active parameter hybrid Mamba-Transformer MoE model released by NVIDIA on August 11, 2026.

NVIDIA has released Nemotron 3.5 Lightning as part of an ongoing effort to refine models for agentic execution environments. Earlier entries in the series provided general capabilities but required additional scaling to reach acceptable performance on tasks involving terminal operations and real-world workflow valuation. The latest model incorporates architectural changes that allow it to deliver stronger results on those specific benchmarks while keeping active parameter usage low during inference.

What prior models set the stage for this release?

Nemotron 3 Nano served as the immediate predecessor and recorded a 7% score on Terminal-Bench v2.1. That result established a lower baseline for compact models in agentic coding and terminal use scenarios. Nemotron 3 Super followed with improved general metrics yet still trailed on the GDPval-AA v2 agentic real-world work tasks benchmark. These earlier releases highlighted the need for targeted gains in efficiency without proportional increases in total parameter count.

The progression across the Nemotron line reflects a deliberate shift toward models suited for repeated invocation in long-running agent systems. NVIDIA identified that high-volume pipelines benefit more from consistent speed than from marginal gains in broad intelligence indices. This focus guided the design choices in the 3.5 Lightning variant.

What benchmark scores define the model's performance?

Artificial Analysis reports a Terminal-Bench v2.1 score of 24% for Nemotron 3.5 Lightning. The figure more than triples the 7% achieved by Nemotron 3 Nano and approaches the level posted by gpt-oss-120b on the same agentic coding and terminal use evaluation. The score derives from the model's ability to handle sequences of commands and file operations within simulated terminal environments.

The model also records a GDPval-AA v2 Elo rating of 824. This rating exceeds the marks set by both Nemotron 3 Super and gpt-oss-120b on the agentic real-world work tasks benchmark. The Elo system in GDPval-AA v2 measures relative performance across tasks that simulate professional workflows such as data analysis and report generation.

On the Artificial Analysis Intelligence Index v4.1.1 the model scores 24. This result ties gpt-oss-120b and improves nine points over the Nemotron 3 Nano predecessor. The index aggregates multiple capability dimensions and places the model on the Pareto frontier for its size category.

How does the model compare to other systems on key metrics?

Direct comparisons show Nemotron 3.5 Lightning closing the gap with models that contain substantially more total parameters. gpt-oss-120b serves as the primary reference point because it posts similar Terminal-Bench v2.1 results. The difference lies in active parameter usage during inference, where the NVIDIA model activates far fewer parameters per token.

The NVFP4 variant listed on the Hugging Face model card reaches 23.46% on Terminal-Bench 2.1. This quantized format preserves most of the accuracy while enabling faster throughput in production deployments. GDPval-AA-V2 scores for the model reach 832 and 865 depending on the checkpoint variant.

Benchmark comparison drawn from Artificial Analysis and NVIDIA Hugging Face model card data.

Model	Terminal-Bench v2.1	GDPval-AA v2 Elo	Active Parameters	Release Date
Nemotron 3.5 Lightning	24%	824	3.6B	August 11, 2026
Nemotron 3 Nano	7%	Not reported	Not reported	Prior to 2026
gpt-oss-120b	Approximately 24%	Below 824	Not reported	Prior to 2026
Nemotron 3 Super	Not reported	Below 824	Not reported	Prior to 2026

What architectural features support its efficiency claims?

The hybrid Mamba-Transformer mixture-of-experts structure activates only 3.6 billion parameters out of the 31.6 billion total during each forward pass. This selective activation reduces memory bandwidth requirements and allows higher token throughput on existing hardware. The design retains capacity for complex reasoning while limiting compute cost per inference step.

NVIDIA provides both NVFP4 and BF16 checkpoints along with speculative decoding optimizations. These elements combine to deliver up to 4x gains in output speed according to the company technical blog. The 1 million token context window further supports extended agent sessions that involve large codebases or lengthy conversation histories.

The model ships under the OpenMDW-1.1 license. This licensing choice facilitates commercial use and modification while maintaining the attribution requirements specified in the license text. Deployment instructions on the Hugging Face page include guidance for integrating the checkpoints into existing inference stacks.

What market and stakeholder implications follow from the release?

The positioning for high-volume execution layers in long-running agentic workflows addresses a practical gap in current deployments. Many organizations run agents that invoke models thousands of times per hour. In such settings the speed and efficiency advantages of Nemotron 3.5 Lightning translate directly into lower operational costs and higher task completion rates.

Integration with NeMo Switchyard enables dynamic routing between this model and larger frontier systems. Enterprises can direct routine terminal and workflow tasks to the compact model while reserving heavier models for edge cases. The NVIDIA blog describes this routing capability as a core element of the release strategy.

Its GDPval-AA v2 Elo of 824 surpasses both Nemotron 3 Super and gpt-oss-120b, while its Terminal-Bench v2.1 score of 24% is >3x Nemotron 3 Nano's 7% and almost matches gpt-oss-120b. For a model of this size and speed, this makes Lightning an attractive model for agentic pipelines.Artificial Analysis, Independent AI benchmarker

What reactions have emerged from independent analysts?

Artificial Analysis highlights the balance of size, speed, and benchmark performance as the primary strength. The independent benchmarker notes that the scores make the model suitable for agentic pipelines where latency and throughput matter more than absolute frontier scores. This assessment appears in the detailed analysis published on the LinkedIn platform.

What steps may follow in subsequent development?

NVIDIA is expected to refine routing logic within NeMo Switchyard to improve task allocation accuracy. Additional checkpoints and quantization formats may appear to further expand deployment options across hardware generations.

1. Deploy the NVFP4 checkpoint first to capture the reported speed improvements in existing pipelines.
2. Configure NeMo Switchyard routing rules to direct terminal and GDPval-style tasks to Nemotron 3.5 Lightning while reserving larger models for open-ended queries.
3. Run internal evaluations on company-specific agent workflows to quantify throughput gains relative to prior Nemotron variants.
4. Review the OpenMDW-1.1 license terms against internal compliance requirements before scaling to production volumes.
5. Monitor updates to the Hugging Face model card for any additional benchmark data or optimization guidance.

The release establishes a template for future compact models that target specialized agent roles rather than general intelligence leadership. Continued iteration on the hybrid architecture could yield additional efficiency gains while preserving the benchmark parity observed in the current version.