NVIDIA has released Nemotron 3.5 Super VL, an open vision-language model built for always-on enterprise agents. Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts (MoE) model from NVIDIA that activates 12B parameters per token. It reads text, images, documents, and video, then reasons, calls tools, and writes code in one model. For developers, one open checkpoint replaces a separate text model along with a separate vision model. NVIDIA also shipped two edge models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B.
TL;DR
- Size: 120B total parameters, 12B active (MoE). Context up to 1M tokens.
- Runs on: 1x DGX Spark, 1x H200, 2x H100, or 1x GB200 with the NVFP4 checkpoint. A BF16 checkpoint is also published.
- Performance: Strong reasoning, agentic coding, and document vision. NVIDIA claims up to 5x higher single-GPU throughput than open peers.
- Best: 89.90 on GPQA-Diamond (NVFP4).
- Worst: 17.78 on tau3-bench banking (NVFP4).
What Is Nemotron 3.5 Super VL?
Nemotron 3.5 Super VL is an open multimodal agent model designed for enterprise workloads that require continuous operation. Its mixture-of-experts architecture activates only 12B of its 120B parameters per token, which keeps inference efficient while preserving large-model capability. The model handles text, images, documents, and video inputs, and it performs reasoning, tool calling, and code generation within a single checkpoint.
By consolidating text and vision capabilities into one model, NVIDIA removes the need for developers to maintain separate text and vision pipelines. The company also introduced two edge-oriented models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B, extending the family to deployment scenarios with tighter compute budgets.
Key Specifications
- Architecture: Mixture-of-experts (MoE)
- Total parameters: 120B
- Active parameters per token: 12B
- Context length: Up to 1M tokens
- Modalities: Text, images, documents, video
- Capabilities: Reasoning, tool calling, code generation
- Checkpoints: NVFP4 and BF16
Deployment and Hardware
The NVFP4 checkpoint is optimized for efficient single-GPU and multi-GPU deployment. According to NVIDIA, it can run on 1x DGX Spark, 1x H200, 2x H100, or 1x GB200. A BF16 checkpoint is also published for teams that require higher-precision weights or different hardware configurations.
NVIDIA claims up to 5x higher single-GPU throughput than open peers when running the NVFP4 checkpoint, positioning the model for always-on enterprise agent deployments rather than occasional or batch inference.
Benchmark Performance
Nemotron 3.5 Super VL shows strong results in reasoning, agentic coding, and document vision. On GPQA-Diamond, the model scores 89.90 under NVFP4. Its weakest reported result is 17.78 on tau3-bench banking under NVFP4, indicating that domain-specific agentic tasks in financial services remain a challenge.
Why It Matters
For developers building enterprise agents, Nemotron 3.5 Super VL offers a unified open checkpoint that replaces separate text and vision models. That simplification reduces integration complexity and infrastructure overhead. Combined with MoE efficiency, a 1M-token context window, and support for tool calling and code generation, the model targets production use cases where agents must continuously perceive, reason, and act across mixed media inputs.
Availability
Nemotron 3.5 Super VL is available as open checkpoints in NVFP4 and BF16 formats. The accompanying edge models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B, are also released as part of the same family.
via MarkTechPost