NVIDIA Nemotron 3.5 Super VL: Specs, License and Hardware

Key Specifications

Developer	NVIDIA (United States)
Release Date	October 15, 2026 (Hugging Face and build.nvidia.com)
Architecture	Hybrid mixture-of-experts with interleaved Mamba-2, MoE and attention layers; Multi-Token Prediction; C-RADIO vision encoder
Parameters	120 billion total / 12 billion active per token
Context Window	Up to 1M tokens per the model card; served context depends on the provider (OpenRouter mentions 300K for some deployments)
Modalities	Text, image and video input
License	OpenMDW License Agreement 1.1 (Linux Foundation, permissive, commercial use allowed)
Checkpoints	NVFP4 and BF16 on Hugging Face; NVIDIA NIM container (vLLM)
API Price Range	About $0.10–0.25 per 1M input tokens and $0.40–0.60 per 1M output tokens, depending on provider
AA Index v4.3	26 (Artificial Analysis Intelligence Index v4.3.x)
Faraday Tier	Starter (1× Mac Studio M5 Max, 128GB) at 4-bit

API Pricing

API prices vary by provider. OpenRouter has listed $0.10–0.15 per million input tokens and $0.40–0.50 per million output tokens, along with a free variant, while Artificial Analysis puts the median across providers at $0.23 and $0.57. The context window you get also depends on who serves the model.

Benchmarks

Most published scores come from NVIDIA’s own runs; the one independent measure is the Artificial Analysis Intelligence Index.

As of October 19, 2026. Vendor-reported scores come from the vendor’s own runs and harnesses, so they aren’t strictly comparable with other labs’ numbers.

On SWE-bench Pro, NVIDIA’s 51.78 is ahead of its own Nemotron 3 Ultra (49.66) but behind Qwen3.8 Flash-Next (60.37), both also NVIDIA-reported. NVIDIA also claims up to 5× the throughput of open models in its size class, and MarkTechPost reports over 600 output tokens per second on Artificial Analysis agentic tasks through a Baseten endpoint. Both come from data-centre GPUs, not Macs, and depend on the workload.

The independent picture is more modest. Artificial Analysis scores it 26 on Intelligence Index v4.3, well above the median of 8 it notes for open-weight models of similar size, but well below the larger open models Faraday lists: GLM-5.3-Flash at 42, MiMo-V2.6-Flash at 38 and MiniMax M3 at 29. It costs roughly $0.20–0.35 per Index task, depending on provider pricing, and generated about 170M tokens on the Index against a 100M median. Scores from v4.1 use a different scale.

How It Compares by Size and License

AA Index is the Artificial Analysis Intelligence Index v4.3.x. “—” means we don’t have a reliable v4.3 score. Sizes for the other models are roughly 0.55 bytes per parameter at 4-bit. Unsloth’s 1- to 2-bit builds of GLM-5.3-Flash also fit Starter, with some loss of quality.

Nemotron 3.5 Super VL is more than four times the size of Qwen3.8-27B and still fits the same single unit. GLM-5.3-Flash and MiMo-V2.6-Flash score higher on the independent index but need roughly twice the memory at 4-bit. Nemotron’s strengths are its size, vision and document work, and its license, not top general intelligence.

Quantized Versions

Which build you run depends on your hardware and how much quality you’ll trade for memory.

Sizes from Unsloth’s guide and Hugging Face repository, community MLX reports and community reports for NVFP4. “—” means our sources give no recommendation. Starter is 128 GB, Growth 256 GB and Scale 384 GB of unified memory.

NVFP4 is NVIDIA’s 4-bit format for its Blackwell GPUs, so on a Mac you use a community MLX build or one of Unsloth’s GGUF builds instead. The GGUF builds run in llama.cpp, and LM Studio lists the model. The 69.5 GB MLX build is text-only, so it drops vision.

File size is not the same as the memory a model needs. Unsloth notes that the operating system, the KV cache (working memory that grows with the conversation), the vision projector and long contexts or video all need room on top of the file. That is why Unsloth recommends 128 GB for an 86.9 GB 4-bit file, and why the 111 GB and 121 GB builds are too tight for a 128 GB machine. Our rule: a model fits a tier only when the whole model sits in unified memory with headroom for working memory.

Lower precision also costs some quality. NVIDIA’s own model card shows a GraphWalks long-context score of 67.69 for the NVFP4 checkpoint against 89.71 for BF16, so a 4-bit build can lose ground on long-context work. That figure covers NVIDIA’s checkpoint, not the MLX or GGUF builds, but lower-bit builds generally trade quality most on long or precise work, which is also why the 1-bit build makes little sense when 4-bit fits.

In tier terms: Starter runs the model at 4-bit, Growth runs an 8-bit build that is closer to full quality, and Scale runs full BF16.

Running It on Faraday

Faraday Machines installs open-weight models on Mac Studio M5 Max hardware in Toronto. Our Starter tier is one Mac Studio with 128 GB of unified memory ($9,999 US), Growth is two units with 256 GB and Scale is three units with 384 GB.

Nemotron 3.5 Super VL fits a single unit, the Starter tier, at 4-bit. Next to Qwen3.8-27B, the small, quick option at about 17 GB, it brings a much larger model to images, documents and agent tasks on the same hardware.

Larger tiers can run higher-precision builds: an 8-bit build on Growth and full BF16 on Scale. That matters most for long-context work. The larger models use most of a tier’s memory, so they run one at a time rather than side by side. On Growth, for example, the 8-bit Nemotron build (132 GB) and GLM-5.3-Flash (roughly 180 GB at 4-bit) can’t both fit in 256 GB.

We don’t quote Mac speeds for this model; the vendor throughput figures above don’t carry over to a Mac Studio. For general reasoning, GLM-5.3-Flash on Growth scores higher on the independent index; see pricing for what each tier includes.

License

NVIDIA’s Nemotron page lists the OpenMDW License Agreement, version 1.1, for all Nemotron 3.5 models. OpenMDW is a permissive license for AI models from the Linux Foundation, developed with Amazon, Meta, IBM and Microsoft, and version 1.1 was released on May 28, 2026. It covers weights, data, code and documentation under one license. Commercial use and derivative models are allowed, and you don’t have to publish your modifications or fine-tuned weights. It was submitted to the Open Source Initiative in mid-August 2026; approval hasn’t been granted yet.

That is simpler than several custom licenses in our license tracker. Kimi K3 uses a custom license with thresholds for very large commercial deployments. GLM-5.3 requires model-as-a-service businesses above $10 billion in annual revenue to pass Z.ai’s security review. Muse Spark Open’s license carries a “Built with” display duty, and our sources show no such duty for OpenMDW. Still, read the license text for notice and redistribution terms before you ship anything built on the model. This is not legal advice.

References