Not established from the available sources.
NVIDIA · MODEL RELEASE TRACKER
Nemotron 3.5 Super VL
Nemotron 3.5 Super VL is an open, customizable multimodal model from NVIDIA described as delivering frontier-level accuracy on scientific benchmarks and fast time-to-solution for scientific coding; the article describes post-training recipes that improved accuracy in a drug-metabolism example and mentions a 16,384-token output budget.CURRENT SNAPSHOT4/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
RELEASE TIMELINE
Published, source-backed release events only.
Nvidia unveils Nemotron 3.5 Super VL and lightweight Bolt models for enterprise agents and local AI PCs
Nvidia announced new open-weight models aimed at always-on agents and edge/PC deployment: Nemotron 3.5 Super VL (120B parameters) for enterprise agents, plus two lightweight Bolt models including Nemotron 3.5 Bolt 74B-A9B 24GB and the smaller Bolt E4B for local AI PCs and workstations. Nvidia says Super VL and Bolt 74B-A9B are available now (Bolt E4B coming later this month), claims Super VL can match larger Nemotron 3 Ultra on instruction-following and outperform some competing models in post-trained agent tasks, and named early users such as Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce.
NVIDIA releases Nemotron 3.5 Super VL, a 120B-parameter open multimodal model for agents
NVIDIA announced Nemotron 3.5 Super VL, an open 120B-parameter mixture-of-experts multimodal model that activates 12B parameters per token to support agentic workflows combining reasoning, instruction following, tool use, and text, image, document, and video understanding. NVIDIA says it achieves up to 5× higher single‑GPU throughput on agentic tasks versus comparable open models, supports post‑training via NeMo (SFT, LoRA, RL) with resources on Hugging Face and a GitHub usage cookbook, and employs inference optimizations such as LiLiCorr speculative decoding and NVFP4 quantization; partners including Scale AI, Ramp, and AI21 report substantial cost, latency, or accuracy improvements in production workloads.
VERIFIABLE FACTS
Every value stays attached to a source and date.
Described as an open, customizable multimodal model (supports multiple input modalities).
Combines reasoning, instruction following, and tool use for agentic workflows.
120B-parameter mixture-of-experts (MoE) architecture that activates 12B parameters per token.
On a single GPU, delivers up to 5× higher throughput for agentic tasks compared to other open models in its size category (as reported by NVIDIA).
Organizations that post-trained the model reportedly achieved improvements such as ~40× lower inference cost, ~10× lower latency, and ~40% higher accuracy for specialized workflows.
Described as an 'open' model (implying open availability or openness to customization).
Uses inference optimizations such as LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse to improve inference efficiency.
Open multimodal model distilled from Nemotron 3 Ultra; NVIDIA publishes the weights, training datasets, and post-training recipes.
120-billion-parameter mixture-of-experts (MoE) architecture that activates 12 billion parameters per token.
Designed for agentic workflows: combines reasoning, instruction following, and tool use to support continuous (always-on) enterprise agents.
On a single GPU, Nemotron 3.5 Super VL delivers up to 5× higher throughput for agentic tasks compared to other open models in its size category (as claimed by NVIDIA).
Announced/launched by NVIDIA in October 2026 (reported 2026-10-19) as part of a set of new open-weight models.
120 billion parameters.
Positioned for use by enterprise agents (always-on agents and edge computing scenarios).
Nvidia says Super VL completes tasks faster than certain competing models (Anthropic and OpenAI models named Opus 5, GPT 6.1 Sol, Luna, Astra) when post-trained on specific agentic tasks such as document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience.
Nvidia listed several companies reported as using Super VL, including Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.
Described as a large model with 120 billion parameters.
Uses fewer parameters than Nemotron 3 Ultra; NVIDIA states it can deliver similar accuracy in instruction-following and tool-use tasks despite the smaller size.
NVIDIA claims Super VL completes tasks faster than certain competing models (named examples include Anthropic models and OpenAI models such as Opus 5 and GPT 6.1 Sol) when post-trained on specific agentic tasks.
120B-parameter Mixture-of-Experts (MoE) model; sparse routing activates 12B parameters per token.
Hybrid Mamba-2, MoE, and attention layers with a RADIO vision encoder; a recursively shared Multi-Token Prediction (MTP) head supports native speculative decoding.
Example usage mentions a 16,384-token output budget for a drug-metabolism prediction workflow.
Targeted use cases include coding (test/validate code, analyze errors, recommend fixes), web research (cross-reference/verify sources), document intelligence (read tables, charts, page layouts), cybersecurity (identify vulnerabilities, triage threats), and video understanding (search, content review, incident analysis). Integrates with major open agent frameworks and orchestration systems via NVIDIA reference implementations and the NVIDIA Agent Toolkit.
NVIDIA reports 63.44% on Hermes Bench, 87.05% on LangChain agent evaluations, and 80.8% on the CharXiv chart-reasoning benchmark.
Reported to deliver 'frontier-level' accuracy on scientific benchmarks.
Reported to provide the fastest time to solution for scientific coding (as claimed by NVIDIA).
In a drug-metabolism prediction example, supervised fine-tuning improved accuracy by 12%, and reinforcement-learning post-training improved it a further 17% (as reported).
WHAT CHANGED