NVIDIA Nemotron 3.5 Super VL Review: Open-Weight Features and Agentic Performance

Overview and Core Capabilities

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. The model contains 120 billion total parameters with 12 billion active parameters in a hybrid Mamba-2 and Transformer mixture-of-experts architecture. It uses the C-RADIOv4 vision encoder and accepts text, images, documents, GUI screenshots, and video clips up to roughly two minutes long. Output remains text only, while the context window reaches one million tokens.

Developers can enable tool use, function calling in JSON format, and multi-token prediction for faster decoding. A toggle in the chat template controls reasoning depth. The model ships under the OpenMDW-1.1 license, which permits commercial use, fine-tuning, and redistribution. NVIDIA supplies BF16 and NVFP4 checkpoints on Hugging Face along with day-zero support in vLLM and NVIDIA NIM.

Reported single-GPU throughput reaches several hundred to more than 600 output tokens per second on H200 or GB200 hardware when running the NVFP4 version. Two H100 GPUs suffice for the quantized checkpoint. Post-training through NVIDIA NeMo supports supervised fine-tuning, LoRA, and reinforcement learning on open datasets.

Pricing, Deployment Options and Comparisons

Weights download at no cost. Hosted inference pricing varies by provider. Typical rates per million tokens fall between $0.10 and $0.25 for input and $0.40 to $1.00 for output, with cached input as low as $0.02 on some endpoints. DeepInfra, OpenRouter, Baseten, and Together AI currently list the model.

Organizations such as Scale AI and EY report accuracy gains of 9 to 44 percent after domain-specific fine-tuning, together with up to 40 times lower inference cost and 10 times lower latency compared with base closed models on specialized tasks. The model tops the Artificial Analysis Intelligence Index among open-weight entries at approximately 26 to 26.5 points, ahead of the prior Nemotron 3 Ultra at 23.

Independent verification of the highest benchmark scores remains limited at the time of writing. Users can add the separate 4B Nemotron 3.5 Content Safety guard model for custom policy enforcement and reasoning traces.

Limitations, Safety Measures and Final Assessment

Training incorporated the Nemotron Content Safety Dataset v2 to filter harmful material including CSAM. No major safety incidents have been reported to date. On-premise deployment gives enterprises direct control over data flows. Video input length caps at roughly two minutes, and performance on long-form video understanding still requires further external testing.

Common user complaints on forums center on the need for substantial GPU memory even in the quantized form and occasional instability during extended multi-turn agent sessions. Several partners note that gains appear most clearly after targeted post-training rather than in the base checkpoint. Independent benchmarks on SWE-bench Verified, BrowseComp, and CharXiv show strong results, yet real-world agent reliability data continues to accumulate.

The combination of open weights, large context, and documented throughput improvements positions the model as a practical option for organizations building specialized multimodal agents, provided they allocate resources for fine-tuning and guardrail integration.