NVIDIA Ships Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter open vision-language model that activates only 12B parameters per token and processes text, images, documents and video in a single checkpoint. The model, built for what NVIDIA calls always-on enterprise agents, reads its inputs, reasons, calls tools and writes code without developers chaining a separate text model to a separate vision model. NVIDIA announced the release on October 15 and shipped two companion edge models alongside it: Nemotron 3.5 Bolt 74B-A9B and Bolt E4B.

The release matters because it compresses a multi-model agent stack into one deployable artifact at a size that fits on a single accelerator. NVIDIA claims up to 5x higher single-GPU throughput than open peers, and the model scores 89.90 on GPQA-Diamond in its NVFP4 configuration — competitive with far larger systems while running on hardware as small as one DGX Spark.

What shipped, and under what license?

Nemotron 3.5 Super VL carries 120B total parameters in a mixture-of-experts layout, activates 12B per token, and supports a context window of up to 1 million tokens. NVIDIA published the weights under the OpenMDW-1.1 license, along with open datasets, training recipes and reinforcement learning environments in NeMo Gym.

The model card lists text, image and video inputs with text output. Video input is limited to MP4 clips of up to 2 minutes. Reasoning can be toggled on or off through the chat template, which lets developers trade latency against accuracy per request.

The NVFP4 checkpoint runs on 1x DGX Spark, 1x H200, 2x H100 or 1x GB200. NVIDIA also published a BF16 checkpoint for deployments that need full-precision weights.

How does the architecture work?

The backbone interleaves Mamba-2 layers, mixture-of-experts layers and attention layers. Each MoE layer holds 512 routed experts. A C-RADIO vision encoder and a connector feed images, document pages and video frames into the language backbone. Those vision components stay in BF16 after quantization, according to the model card.

NVIDIA attributes the model's speed to four techniques:

• Multi-token prediction: a native drafter proposes several tokens, verified in one forward pass.
• LiLiCorr: a speculative drafter for DGX Spark that delivers up to 2.3x faster token generation.
• NVFP4 with quantization-aware distillation: 4-bit weights and activations on Blackwell, 4-bit weights with 16-bit activations elsewhere.
• Context reuse: prefix caching keeps both attention caches and Mamba states across agent turns.

NVIDIA distilled the model from Nemotron 3 Ultra and states it trained on about 26 trillion tokens. Post-training used roughly 3.8 million samples plus reinforcement learning across dozens of environments. Continued pretraining added 10 billion Indic tokens.

How does it perform on benchmarks?

All scores below are NVIDIA-reported, from the NVFP4 checkpoint on the model card.

Area	Benchmark	Score
Reasoning	GPQA-Diamond	89.90
Coding	SWE-bench Verified	71.07
Coding	Terminal-Bench 2.1	60.80
Web agent	BrowseComp	77.5
Instruction following	IFBench	71.50
Charts	CharXiv reasoning	79.88
Documents	MMLongBench-Doc	62.82
Computer use	OSWorld-Verified	52.3
Long-horizon agent	tau3-bench banking	17.78

The spread between the best and worst scores is the story in this table. The model handles graduate-level reasoning at 89.90 on GPQA-Diamond but manages only 17.78 on tau3-bench banking, a long-horizon agentic benchmark. Enterprise buyers evaluating the model for multi-step financial workflows will need to weigh that gap against its stronger document and coding results.

On agent harness evaluations, Super VL beats Nemotron 3 Ultra, the model it was distilled from. It scores 63.44% on Hermes Bench versus 56.33% for Ultra, and 87.05% on LangChain evaluations versus 84.3%. Ultra carries about 4.6x more total parameters.

On throughput, NVIDIA reports over 600 output tokens per second on Artificial Analysis agentic tasks through a Baseten endpoint, and its chart marks a 5x speed gap over Inkling Small.

How does it compare with similar open models?

Feature	Nemotron 3.5 Super VL	Qwen3.5-122B-A10B	Inkling-Small	Nemotron 3 Ultra
Developer	NVIDIA	Alibaba Qwen	Thinking Machines Lab	NVIDIA
Total / active params	120B / 12B	122B / 10B	276B / 12B	550B / 55B
Context	Up to 1M	262,144 native, extensible to 1,010,000	Not disclosed	Evaluated at 1M (Ruler @1M)
Inputs	Text, image, video	Text, image, video	Text, image, audio	Not disclosed
License	OpenMDW-1.1	Apache 2.0	Apache 2.0	OpenMDW-1.1
Reference hardware	1x DGX Spark, 1x H200, 2x H100 or 1x GB200	8-GPU tensor-parallel serving example	Not disclosed	Not disclosed
GPQA Diamond	89.90	86.6	89.5	Not disclosed
SWE-bench Verified	71.07	72.0	80.2 (bash-only harness)	65 to 70.4 (across 5 harnesses)
IFBench	71.50	76.1	82.2	82
CharXiv reasoning	79.88	77.2	77.4	Not disclosed
BrowseComp	77.5	63.8	77.4 (with context management)	Not disclosed
Terminal-Bench	60.80 (v2.1)	49.4 (v2)	64.7 (v2.1, best harness)	54 (v2.0)

The scores are self-reported by each vendor, and harnesses and benchmark versions differ. Gaps of a few points carry limited weight. The clearest differentiation is not in the score table but in the deployment column: no listed peer matches a reference configuration that starts at a single workstation-class GPU.

How can developers deploy it?

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

For production serving, NVIDIA's Dynamo recipes add KV-aware routing. On 16 B200 GPUs at 50 tokens per second per user, NVIDIA reports the stack serves 8x more users, with first-token latency dropping 85%.

Customization runs through NeMo AutoModel for supervised fine-tuning and LoRA, and NeMo RL for reinforcement learning. NVIDIA cites three early customer results:

• Scale AI, for Skydance vendor credit processing: 40x lower inference cost and 5x lower latency.
• Ramp, for receipt-to-JSON extraction: up to 50% higher throughput and 10x lower latency expected.
• AI21, for the Alice threat intelligence product: 40% higher accuracy.

What are the Nemotron 3.5 Bolt models?

Nemotron 3.5 Bolt 74B-A9B targets local agents. NVIDIA distilled it from Super VL, and native 2-bit quantization fits the model into 24GB of memory on RTX and DGX Spark systems. NVIDIA says Bolt ranks first among open models in its class on PinchBench and second on BrowseComp among open models. It is available now on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

Bolt E4B, a 4GB model, arrives October 23. It targets voice assistants and embedded agents on GeForce RTX, Jetson, and RTX or DGX Spark hardware.

What are the key takeaways?

• 120B MoE with 12B active parameters, 1M-token context, open weights under OpenMDW-1.1.
• One model handles text, images, documents, video, tool calls and code generation.
• The NVFP4 checkpoint runs on a single DGX Spark, H200 or GB200.
• 89.90 on GPQA-Diamond, but 17.78 on tau3-bench banking.
• Bolt 74B-A9B brings a local agent model to 24GB RTX-class PCs, with Bolt E4B following October 23.

NVIDIA has also published a technical report, the Hugging Face model card and a usage cookbook for the model. Together with the open datasets and RL environments in NeMo Gym, the release gives developers the full stack — weights, recipes and evaluation environments — to adapt the model rather than merely run it.