What it is
Nemotron 3.5 Super VL is NVIDIA's flagship open vision-language model, released on 15 October 2026. It is a mixture-of-experts model with 120 billion parameters, 12 billion active per token, that mixes Mamba-2 layers (a fast alternative to attention for long sequences) with standard attention, adds a C-RADIO image encoder and has a built-in multi-token prediction head for speculative decoding. It reads text, images, documents and mp4 video up to two minutes long, writes text and accepts up to 1M tokens of context. Reasoning can be off, medium or full.
Full-precision weights need four H100s or two GB200s; a 4-bit NVFP4 version, about 75 GB on Hugging Face, targets Blackwell GPUs and DGX Spark. NVIDIA publishes training datasets and recipes alongside the weights, though the card lists many sources as private or vendor-supplied. The licence is OpenMDW-1.1, which allows commercial use.
Two smaller Bolt models came with it: Bolt 74B-A9B, compressed with 2-bit weights to fit in 24GB with a 256K context, and Bolt E4B, a 4GB voice-agent model due on 23 October. Super VL is served by NVIDIA NIM, by OpenRouter from $0.15 input and $0.50 output per million tokens with a free variant, and by eight API providers in Artificial Analysis's count.
Our review
Super VL is the strongest open model in its size class on independent measurement, and its weak spot is one NVIDIA prints itself. Artificial Analysis scores it 26 on its Intelligence Index, first of 67 open-weights models between 40B and 150B parameters, where the median is 8 and Mistral Small 4 scores 11. It also measured about 450 output tokens a second across providers, the fastest of 26 in the class, with 0.72 seconds to first token.
Coding is the exception. NVIDIA's own table puts it at 51.78 on SWE-bench Pro against 60.37 for Qwen3.8-Flash-Next and 59.70 for GLM-5.3-Flash, and 60.67 on Terminal-Bench 2.1 against 85.80 and 82.70. Its wins are restraint and long context: a 30.0 hallucination rate on AA-Omniscience against 41.3 for Qwen and 95.9 for DeepSeek-V4-Flash, and 89.71 on GraphWalks, the best in the table. On documents it is close to rivals rather than ahead, with 62.82 on MMLongBench-Doc to Qwen's 67.9.
The openness is real but partial. Datasets and recipes are released and linked from the Hugging Face card, which matters to buyers who must document how a model was built, yet the card also lists many training sources as private or undisclosed. OpenMDW-1.1 allows commercial use without the attribution and revenue clauses of MiniMax's licence.
It is also verbose and very NVIDIA-shaped. Artificial Analysis counted 170 million tokens generated across its index against a 100 million median, which eats into the low $0.15/$0.50 OpenRouter price, and the fast path runs through Hopper or Blackwell GPUs, NVFP4, NIM containers and a pinned vLLM or SGLang build. The card's quick-start and its deployment table even disagree on the number of speculative tokens, three against five, a small sign of a release still settling.
Where it earns its place
A research-platform team at a pharmaceutical company
Answer questions across long papers, charts and lab reports
Feed papers with figures into the long context with reasoning on. The low hallucination score and released recipes help an internal model-risk review, and on OpenRouter 50,000 queries a month at 100K input and 3K output tokens each costs about $825.
An ML platform engineer with an H100 or Blackwell cluster
Self-host a multimodal agent model without per-token fees
Serve BF16 across four H100s or two GB200s, or the 75 GB NVFP4 build on Blackwell, with SGLang's EAGLE speculative decoding. One node then runs document, image and short-video jobs in-house under a licence that needs no notice to NVIDIA.
A compliance lead at a European insurer
Pick a model whose provenance can be shown to auditors
Point auditors at the published pre- and post-training datasets and recipes, noting that some sources are private. OpenRouter offers no EU in-region routing for this model, so EU data residency means self-hosting or a provider with EU hosting.
Before you commit
- Setup
- Hosted: call nvidia/nemotron-3.5-super-vl on OpenRouter or use NVIDIA NIM. Self-hosted: vLLM v0.32.0 or NVIDIA's SGLang image with the nemotron_v3 reasoning parser; thinking is switched through the chat template (enable_thinking, medium_effort).
- Integrations
- Hugging Face, ModelScope, NVIDIA NIM, OpenRouter and other API providers; vLLM, SGLang and Transformers; tool calling through the qwen3_coder parser.
- Data and privacy
- OpenMDW-1.1 permits commercial use. Self-hosting keeps data in-house; hosted providers set their own retention terms, and OpenRouter does not offer EU in-region routing for this model.
- Cost at scale
- On OpenRouter's $0.15/$0.50, 1M agent turns a month at 8K input and 1K output tokens comes to about $1,700, before its verbosity inflates output. Self-hosting at BF16 means four H100s per replica; NVFP4 fits a smaller Blackwell footprint.
- Limits
- 1M-token context on the card, but OpenRouter's top provider serves 262K. One image and one video of up to two minutes per request. Reasoning tokens count against max_tokens. Coding and terminal-agent scores trail Qwen3.8-Flash-Next and GLM-5.3-Flash.
What's good
- First of 67 mid-size open models on Artificial Analysis's index
- About 450 tokens/s across providers, per Artificial Analysis
- Low hallucination rate (30.0) in NVIDIA's table
- Open weights, datasets and recipes under a commercial licence
What's not
- Trails Qwen3.8-Flash-Next and GLM-5.3-Flash on coding in NVIDIA's own table
- Verbose: 170M tokens on AA's index against a 100M median
- BF16 needs four H100s; NVFP4 needs Blackwell
- One image and one two-minute video per request
Best for
Teams running document and research agents who want the strongest open mid-size model, especially on NVIDIA hardware.
Alternatives
Nemotron 3.5 Super VL: common questions
- Is Nemotron 3.5 Super VL worth it?
- The open model to beat at 120B on independent tests, provided the job is documents and research rather than code. We rate it 4.1 out of 5.
- How much does Nemotron 3.5 Super VL cost?
- Free weights under OpenMDW-1.1. Hosted: OpenRouter from $0.15 input / $0.50 output per 1M tokens, plus a free variant; NVIDIA NIM and other providers priced separately.
- What are the downsides of Nemotron 3.5 Super VL?
- Trails Qwen3.8-Flash-Next and GLM-5.3-Flash on coding in NVIDIA's own table; verbose: 170M tokens on AA's index against a 100M median; bF16 needs four H100s; NVFP4 needs Blackwell; one image and one two-minute video per request.
- Who is Nemotron 3.5 Super VL best for?
- Teams running document and research agents who want the strongest open mid-size model, especially on NVIDIA hardware.
- How much does Nemotron 3.5 Super VL cost for a team or in production?
- On OpenRouter's $0.15/$0.50, 1M agent turns a month at 8K input and 1K output tokens comes to about $1,700, before its verbosity inflates output. Self-hosting at BF16 means four H100s per replica; NVFP4 fits a smaller Blackwell footprint.
- Is Nemotron 3.5 Super VL safe for company data?
- OpenMDW-1.1 permits commercial use. Self-hosting keeps data in-house; hosted providers set their own retention terms, and OpenRouter does not offer EU in-region routing for this model.
- What are the best Nemotron 3.5 Super VL alternatives?
- The closest alternatives we have reviewed are Qwen3.8-Flash-Next, GLM-5.3-Flash, DeepSeek-V4.1-Flash and gpt-oss-120b.
Answers checked Oct 23, 2026.
Make Nemotron 3.5 Super VL? Show this rating on your site
