Nvidia Opens 120B Nemotron 3.5 Super VL, Claims Task Wins
TL;DR
- Nvidia's 5x throughput claim cites the Artificial Analysis Intelligence Index leaderboard, which had not listed Nemotron 3.5 Super VL as of its launch day.
- Named post-trainers Scale AI and AI21 report 40x inference cost reduction and 41% accuracy gains on proprietary specialized workflows after post-training.
- Super VL completes 4x more enterprise jobs than its Nemotron predecessor and uses 123x fewer tokens per correct task, per Nvidia's own figures.
Nvidia on October 15 released Nemotron 3.5 Super VL, a 120-billion-parameter open multimodal model aimed at agents that read PDFs, charts, tables, images and webpages, and says post-trained versions beat Claude Opus 5 and GPT-6.1 Sol on five domain tasks.
The comparison chart in VentureBeat's writeup covers "document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims," with gains Nvidia puts at 9 to 44% after post-training. On medical insurance claims, accuracy rises from roughly 28% to 72%, compared with about 38% for Claude Opus 5. On security attack resilience it rises from roughly 35% to 76%, compared with about 66% for GPT-6.1 Sol. The report flags that "NVIDIA has not supplied the task definitions, sample sizes or testing protocol with the chart, so a buyer should test the same approach against its own work."
Nvidia is also pushing two smaller Bolt variants: a 74B-A9B model with native 2-bit compression that fits in 24GB of GPU memory, and a 4GB Bolt E4B edge model slated for October 23. The training recipe spans roughly 26 trillion tokens, 3.8 million post-training samples, and reinforcement learning across about 100 open environments. Weights ship under Nvidia's OpenMDW license, which "permits commercial use, fine-tuning and redistribution," and distribute through Hugging Face, OpenRouter, ModelScope and NIM, with a rate-limited free Nvidia-hosted API for development and NVIDIA AI Enterprise at $4,500 per GPU per year for production.
Nvidia describes Super VL as "up to five times faster than a comparable open model," citing near 625 output tokens per second on four Blackwell B200 GPUs. The launch follows Nvidia's agreement to acquire Hugging Face, a major distribution platform for open models.
The drop lands in a run of Nvidia coverage our tracker has clocked at 201 stories in the last 90 days, and in the broader open-source lane where 344 releases have crossed the feed in the same window.
What others are reporting
-
NVIDIA Developer Blog Read →
First-party post details LiLiCorr speculative decoding, NVFP4 quantization distillation, and production results from Scale AI and AI21 with named 40x cost and 41% accuracy figures.
On a single GPU, this model delivers up to 5x higher throughput for agentic tasks compared to other open models in its size category.
-
WCCFTech Read →
Covers the full Nemotron 3.5 stack including Bolt 4GB and 24GB edge models, and traces the predecessor Lightning's orchestration bottleneck as the problem Super VL was built to fix.
The new Super build finishes 4x the share of enterprise jobs its predecessor did and spends 123x fewer tokens per correct task to do it.
-
Tech Insider Read →
Critical framing: public leaderboards still rank Claude Opus 5.5 and Gemini 4 Argon ahead, the 'free' model carries $4,500/GPU/year enterprise costs, and the launch is framed as a geopolitical play against Chinese open-source expansion.
Unlike a routine point release, Super VL ships with a specific pitch: agentic reasoning across PDFs, charts, tables, images, and webpages.
-
Orca Router Read →
Surfaces a Terminal-Bench 2.1 score discrepancy between the model card (66.70) and the technical report (60.67), and notes the cited leaderboard has not yet listed the model.
As of this writing, Artificial Analysis does not list Super VL at all.
Originally reported by venturebeat.com
Read the original article →Original headline: Nvidia Open-Weights Nemotron 3.5 Super VL, a 120B MoE With 12B Active Parameters and 1M Context