Nemotron 3.5 Super VL 120B: Open Multimodal Agent Model

NVIDIA has released Nemotron 3.5 Super VL 120B, an open-weight multimodal agent model designed for enterprise workloads that process text, images, documents and video within a single system. The company positioned the release to help teams building always-on agents avoid the operational overhead of maintaining separate models for language and vision tasks.

What changed

NVIDIA announced general availability of Nemotron 3.5 Super VL, a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token. The model reaches a 1 million token context window and runs on a single DGX Spark, H200, dual H100, or one GB200 GPU. NVIDIA also released two smaller variants: Nemotron 3.5 Bolt 74B-A9B for local agents on 24GB systems, and Nemotron 3.5 Bolt E4B 4GB, arriving October 23 for voice assistants and embedded agents. All three models ship under the OpenMDW-1.1 license with open weights.

Super VL integrates text, image and video understanding in one backbone, enabling agents to process web pages, PDFs, documents and structured data without routing tasks between separate systems. The architecture combines Mamba-2, mixture-of-experts and attention layers, with 512 routed experts per MoE layer. Vision inputs pass through a C-RADIO encoder and connector before entering the language backbone.

NVIDIA trained the model on approximately 26 trillion tokens during pretraining and continued pretraining. Post-training used roughly 3.8 million samples and reinforcement learning across dozens of environments. Continued pretraining also added 10 billion Indic language tokens to expand language coverage.

The company reports four speed techniques: multi-token prediction where a drafter proposes multiple tokens verified in one forward pass, LiLiCorr speculative drafting for DGX Spark enabling up to 2.3x faster token generation, NVFP4 4-bit quantization with quantization-aware distillation, and context reuse through prefix caching of attention and Mamba states across agent turns.

Why it matters

Teams deploying enterprise agents typically manage separate infrastructure for language and vision tasks, adding latency and operational complexity as context moves between models. A unified multimodal model eliminates that friction, letting context flow through reasoning, tool calls and code generation without intermediate routing.

NVIDIA reports Super VL outperforms its own larger Nemotron 3 Ultra model on agentic benchmarks despite carrying approximately one-fifth the total parameters. On Hermes Bench, Super VL scores 63.44% compared to Ultra’s 56.33%. On LangChain evaluations, Super VL achieves 87.05% versus Ultra’s 84.3%. The company reports over 600 output tokens per second on Artificial Analysis agentic tasks, representing faster task completion for workflows spanning days or weeks where per-token latency compounds across thousands of agent steps.

Organizations can deploy Super VL on-premises and customize it on proprietary data. Teams in financial services, insurance and healthcare can run localized versions without sending customer data to cloud APIs. Scale AI adapted Super VL for Skydance’s vendor credit processing and EY’s M&A due diligence workflows. Ramp customized the model for receipt-to-JSON extraction, AI21 and Alice adapted it for cybersecurity threat intelligence, Mercor fine-tuned it for software engineering tasks, and Lila Sciences applied it to scientific discovery. Additional organizations including Dream, R1 and TrainLoop are building specialized versions for their domains.

The Bolt 74B-A9B variant enables local agent development on consumer RTX and Jetson hardware. NVIDIA reports it ranks first among open models in its class on PinchBench and second on BrowseComp. The 4GB Bolt E4B model targets voice assistants and embedded interfaces, extending agentic capabilities to edge devices where model size and latency create hard constraints.

What to test

Before committing Super VL to production deployments, validate these vendor claims on your own workloads:

• Throughput parity: NVIDIA claims 5x higher single-GPU throughput than comparable open models. Test this on your actual agentic tasks and hardware, as benchmark conditions often differ from production load patterns.
• Long-context stability: The 1 million token context is available, but verify whether reasoning quality holds at your typical usage window. Query-document pairs and multi-turn agent traces can saturate context differently.
• Vision accuracy on your documents: NVIDIA emphasizes optical character recognition, tables, charts and structured data handling. If your agents work with handwritten forms, non-English text or specialized diagrams, run samples through the model first.
• Quantization trade-offs: The NVFP4 4-bit checkpoint delivers speed gains but may impact reasoning on edge cases. Compare BF16 and quantized versions on your benchmark set.
• Customization cost: Early adopters report strong results with Scale AI and Ramp, but fine-tuning timelines and data requirements vary across use cases. Test your own post-training pipeline with a small dataset before scaling.
• Benchmark variance: NVIDIA reports 89.90 on GPQA-Diamond but only 17.78 on tau3-bench banking with NVFP4. Understand where the model excels and where it struggles before committing to high-stakes tasks.

Common questions

Can I run Super VL locally?

The 120B full model requires a single high-end GPU (H200, DGX Spark, GB200). For local on-device deployment, NVIDIA offers Bolt 74B-A9B, which uses 2-bit quantization to fit 24GB of memory on RTX and DGX Spark systems. The Bolt E4B 4GB model targets even lighter embedded use cases on GeForce RTX, Jetson and RTX or DGX Spark.

How does Super VL compare to Nemotron 3 Ultra?

NVIDIA reports Super VL outperforms Ultra on agentic benchmarks despite carrying one-fifth the parameters. On Hermes Bench, Super VL scores 63.44% compared to Ultra’s 56.33%. On LangChain evaluations, Super VL achieves 87.05% versus Ultra’s 84.3%. The smaller model prioritizes agentic speed and efficiency over raw reasoning scale.

What makes it faster than other open models?

NVIDIA credits four techniques: multi-token prediction where a native drafter proposes multiple tokens verified in one forward pass, LiLiCorr speculative drafting for DGX Spark enabling up to 2.3x faster token generation, NVFP4 4-bit quantization with quantization-aware distillation, and context reuse through prefix caching of attention and Mamba states across agent turns.

Where can I get the model?

Super VL is available on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM microservice. Supported runtimes include vLLM, SGLang and TensorRT-LLM on Blackwell and Hopper GPUs. Local deployment options include llama.cpp and Unsloth.

The conclusion

Nemotron 3.5 Super VL 120B addresses a genuine gap in open agentic models. The unified multimodal approach reduces operational complexity for teams building always-on agents. The 1 million token context window and claimed efficiency gains matter for long-running workflows. The open weights, datasets and training recipes support the customization that enterprise deployments increasingly require.

The benchmark spread, showing strong reasoning on GPQA-Diamond but weaker results on domain-specific tasks like tau3-bench banking, signals the model is built for generalist agentic work rather than vertical specialists. Teams in finance, legal or healthcare should test on representative documents before production deployment. The three-tier release covering cloud to edge also indicates NVIDIA is deliberately addressing deployment tiers that closed models often overlook.

Watch for early adopter results from Scale AI, Ramp and others to confirm whether post-training truly closes the gap on specialized workflows. If customization timelines remain short and accuracy gains hold, open agentic models may accelerate adoption of on-premises agent infrastructure across regulated industries.

Related reading

• LFM2.5-24B-A2B: Fast on-device reasoning agents
• RTX Spark PC AI launches Oct 16 with local agents
• QwenGyre elastic RL framework trains long-horizon agents