Open Source · OCTOBER 16, 2026
Nemotron 3.5 Super VL lands at 120B MoE, 89.9 on GPQA-Diamond, $0.25 per Intelligence Index task
NVIDIA's open flagship pairs a hybrid vision-language stack with NVFP4 and multi-token prediction to hit ~625 output tokens/sec on four B200s, at roughly 85% below the prior generation's per-task cost.
NVIDIA released Nemotron 3.5 Super VL on October 15, a 120-billion-parameter mixture-of-experts checkpoint shipping under the OpenMDW-1.1 license with 12 billion parameters active per token. The headline figure, per Artificial Analysis, is $0.25 per Intelligence Index task against $1.64 for the prior Nemotron 3 Super: an 85% per-task drop in one generation.
The benchmark card does the kind of work NVIDIA presumably wanted it to do. 89.90 on GPQA-Diamond (NVFP4). 71.07 on SWE-bench Verified. 60.80 on Terminal-Bench 2.1. 77.5 on BrowseComp. 71.50 on IFBench. 62.82 on MMLongBench-Doc. 52.3 on OSWorld-Verified. 79.88 on CharXiv reasoning. On agentic evaluations the model clears 63.44% on Hermes Bench against Nemotron 3 Ultra's 56.33%, and 87.05% on LangChain evaluations against Ultra's 84.3%. Long-horizon reliability remains the soft underbelly: 17.78 on tau3-bench banking is a reminder that multi-step reasoning collapses faster than headline scores suggest.
The Intelligence Index itself is where the pricing story sharpens. Artificial Analysis scores Super VL at 26 (26.5), against 13 for the prior generation and an open-weights median of 8 in the same size class. The firm calls it the top-scoring permissively licensed US open-weight model in its class. It still trails GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40, so "top open US" is doing load-bearing work in that sentence.
Priced endpoints run $0.20 per million input tokens, $0.575 per million output, and $0.055 cached. Serving characteristics are the quieter story: ~625 output tokens per second on four Blackwell B200s at 100,000 input tokens, roughly 5x Thinking Machines Lab's Inkling Small at ~125. A 16-B200 Dynamo configuration with KV-cache-aware routing serves up to 8x more users at 50 tokens/sec each, with first-token latency down 85%. The context window is one million tokens; video input runs to two minutes of MP4.
For an owner-operated business, none of this is a model to run. It's a cost floor. When frontier-class multistep reasoning, document understanding, and tool use settle at $0.575 per million output tokens on self-hosted infrastructure, the economics of done-for-you AI sales and marketing shift from aspirational to defensible. That's the same curve Claude Haiku 5.5's $0.10/M price cut traced on the closed side, and the same one Salesforce's Koa post-trained on Nemotron 120B banked on at the application layer. Services like LemonLime, which prepare prospects, outreach, and content proactively for small businesses at $999/month, inherit this curve directly. The owner who never touches a checkpoint still gets a cheaper, faster backend.
Whether the numbers survive third-party replication is the only question that matters now. NVIDIA ran the index. Everyone else gets to check it.
Sources
- https://venturebeat.com/technology/nvidias-new-open-nemotron-3-5-super-vl-with-post-training-beats-claude-opus-5-gpt-6-1-sol-on-specialized-task-benchmarks
- https://www.hpcwire.com/aiwire/2026/10/15/nvidia-expands-nemotron-family-with-highest-efficiency-super-vl-and-bolt-models/
- https://www.marktechpost.com/2026/10/15/nvidia-nemotron-3-5-super-vl-release/
- https://www.unite.ai/nvidia-releases-open-nemotron-3-5-super-vl-and-bolt-edge-models/
- https://artificialanalysis.ai/models/nemotron-3-5-super-vl