NVIDIA launched Nemotron 3.5 Super VL on October 15, 2026 — a 120-billion-parameter open-weight model built for always-on AI agents that read text, images, documents, and video in one pass. It's the company's strongest open-model play yet, and it landed the same week Meta claimed the US open-model crown with Muse Spark Open.
What's new
The model is a mixture-of-experts system with 120 billion total parameters and 12 billion active per token. It takes text, image, and video input (up to 1M-token context) and produces text output, with reasoning that can be toggled on or off. It was built with the Nemotron Coalition, trained on about 26 trillion tokens with post-training and reinforcement learning across 100 open environments.
The numbers
NVIDIA claims up to 5x higher single-GPU throughput than other open models in its class, and performance comparable to its largest reasoning models at one-fifth the size — company-reported benchmarks, so treat those as claims. Independently, Artificial Analysis calls Super VL the leading US open-weight model, scoring 26.5 on its Intelligence Index — roughly double its predecessor Nemotron 3 Super (13), and ahead of Thinking Machines' Inkling Small (25.7) and NVIDIA's own Nemotron 3 Ultra (23).
Pricing and availability
Free to download. Weights are on Hugging Face, OpenRouter, and ModelScope under the OpenMDW-1.1 license, which permits fine-tuning, redistribution, and commercial use. The NVFP4-quantized checkpoint runs on a single DGX Spark, H200, or GB200 (a BF16 checkpoint is also published). NVIDIA also released open datasets, training recipes, and reinforcement-learning environments — a full reproducibility package. Two smaller edge models launched alongside: Nemotron 3.5 Bolt 74B-A9B (2-bit quantized, fits in 24GB, runs on RTX laptops) and Bolt E4B 4GB (due October 23, aimed at Jetson and edge devices).
Why it matters
NVIDIA is going all-in on open weights right as the US open-model race heats up — Meta's Muse Spark Open landed the same week with the same "best US open model" claim, so expect independent benchmarks to settle this. Enterprise adoption is already lined up: Scale AI, Ramp, and AI21 are post-training it for finance, receipt parsing, and threat intel, and platform partners including Atlassian, CrowdStrike, Salesforce, and Perplexity are integrating it into their agent products.
Verdict
Super VL is NVIDIA's most serious open-model release to date — genuinely open, agent-grade, and already moving into enterprise workflows. Keep an eye on the benchmark fight: with Meta and NVIDIA both claiming the US open crown in the same week, independent tests will decide this round.
Comments
Post a Comment