Nvidia releases Nemotron 3.5 Super VL: open 120B multimodal MoE with 12B active parameters

Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters, of which only 12 billion are active per token. It is built for enterprise agents that run continuously and process text, images, documents, and video. Two smaller models, Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB, target local and edge hardware.

Key specifications

• Architecture: Mixture-of-Experts, 120B total parameters, 12B active per token
• Modalities: text, images, documents, video
• Origin: distilled from Nemotron 3 Ultra, Nvidia's largest reasoning model
• Quantization: 4-bit NVFP4; runs fully in 4-bit on Blackwell GPUs, while older cards store weights at 4 bits and compute at 16 bits
• Context window: not disclosed in the source material
• Pricing: not yet disclosed; weights are open
• Training cutoff: not disclosed

Nvidia developed the model with the Nemotron Coalition, which includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs. It is publishing datasets and training recipes alongside the weights. Customization runs through the NeMo framework, Python pipelines, and open datasets.

Performance claims

All figures below come from Nvidia unless noted.

• Nvidia claims Super VL matches Nemotron 3 Ultra's accuracy on instruction-following, tool-calling, multi-step business process, and image-analysis tasks with about one-fifth the parameters. It reportedly scores higher than Ultra in agent tests from Hermes and LangChain.
• It beats its predecessor, Nemotron 3 Super, on SWE-bench and Terminal-Bench. On one coding benchmark it is slightly less accurate than competitors but finishes tasks four times faster, according to Nvidia.
• On a single GPU, Nvidia claims up to 5x the throughput of similarly sized open models on agent tasks, citing the Artificial Analysis leaderboard. In that overall ranking, Super VL trails leading closed models by a wide margin but beats Nemotron 3 Ultra.
• In an independent test by EXO Labs on the DGX Spark, no other model was both more accurate and faster.
• On CharXiv, Super VL scores 80.8%, versus 63% for Nemotron 3 Nano Omni. With Python execution, it reaches 84%.

Inference efficiency

With Nvidia's Dynamo inference software on 16 B200 GPUs, Nvidia says Super VL serves 8x as many concurrent users and cuts time to first token by 85%. It does this by caching repeated prompt components and routing session requests to the GPU holding that session's cache. A built-in draft component supports speculative decoding. On the DGX Spark, Nvidia recommends its LiLiCorr method, which it says speeds up generation by up to 2.3x.

Bolt models for local hardware

Bolt 74B-A9B 24GB is a hybrid MoE with 9B of 74B parameters active per step. It uses 2-bit weight quantization to fit in 24 GB of memory. It is optimized for RTX Spark and DGX Spark (up to 128 GB shared memory) and also runs on GeForce RTX PCs, DGX Station, and RTX PRO workstations. It supports llama.cpp, Ollama, LM Studio, and vLLM.

Bolt E4B 4GB targets voice assistants, interactive characters, and device interfaces on GeForce RTX, Jetson, and RTX/DGX Spark. It has not shipped yet; Nvidia says it will follow later this month.

Early adopters

Nvidia says dozens of organizations are fine-tuning Super VL. Scale AI built a vendor-credit agent for Skydance, with costs expected to fall to one-fortieth and response time to one-fifth. Ramp uses it to convert receipt photos into structured data and expects 10x faster responses. EY uses it for acquisition due diligence. These results are Nvidia-reported expectations, not independent measurements.

What this means

Super VL is aimed at enterprises that want to fine-tune on proprietary data and self-host. Open weights, datasets, and recipes make that practical, but every variant is optimized for Nvidia hardware, which ties the release to Nvidia's chip and system sales. The efficiency claims (throughput, time to first token, concurrency) are the strongest selling point, and most of them are vendor-reported, with EXO Labs and Artificial Analysis the only external sources cited. The gap to leading closed models on overall intelligence remains wide. The local Bolt 74B model still needs 24 GB of memory, which puts it out of reach for most consumer hardware. Until the 4 GB variant ships, the on-device story is limited to developers and workstation owners.