RELEASE · MODELS · #2046
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.
KEY POINTS
- 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.
- A single open model that combines vision, language, reasoning, and tool use—while improving inference efficiency—can simplify enterprise agent stacks and materially reduce costs and latency for specialized workflows.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter mixture-of-experts (MoE) multimodal model that activates ~12B parameters per token and is intended for agentic workflows combining reasoning, instruction following, tool use, and text/image/document/video understanding. NVIDIA publishes post‑training resources (NeMo workflows, datasets on Hugging Face, a GitHub cookbook) and describes inference optimizations (LiLiCorr speculative decoding, NVFP4 quantization, multi-token prediction, context reuse). NVIDIA and partners present single‑GPU throughput and production improvements versus comparable open models.
WHY NOW
A single open multimodal MoE model with claimed high single‑GPU throughput and post‑training tools could simplify enterprise agent stacks and change cost/latency tradeoffs for long‑running, multi‑step agent workflows—making this release immediately relevant to teams building or optimizing agents.
WHO IS AFFECTED
Developers and teams building enterprise agents, document- and multimodal‑analysis pipelines, and organizations using NVIDIA hardware for inference. NVIDIA ecosystem partners and users of NeMo/Hugging Face resources are directly implicated.
CONFIRMED
• NVIDIA announced Nemotron 3.5 Super VL as an open, customizable multimodal MoE model with 120B parameters and ~12B parameters active per token (NVIDIA). • The model is described as combining reasoning, instruction following, tool use, and text/image/document/video understanding in one model (NVIDIA). • NVIDIA says developers can post‑train the model via NeMo (SFT, LoRA, RL) and that datasets/recipes are published on Hugging Face and a GitHub usage cookbook (NVIDIA). • NVIDIA reports inference optimizations including LiLiCorr speculative decoding and NVFP4 quantization with quantization‑aware distillation, multi‑token prediction, and context reuse (NVIDIA). • NVIDIA states the model delivers up to 5× higher single‑GPU throughput on agentic tasks versus comparable open models of its size (NVIDIA). • NVIDIA cites partner post‑training by organizations (Scale AI, Ramp, AI21) claiming improvements (examples given include 40× lower inference cost, 10× lower latency, 40% higher accuracy) (NVIDIA). • Media reporting notes two smaller Bolt variants for local machines (24GB Bolt and a 4GB Bolt E4B planned) and that Nemotron Super VL was distilled from Nemotron 3 Ultra; independent testing (EXO Labs) on DGX Spark is reported as finding no other model both more accurate and faster in their test (media).
UNCERTAIN
Missing or conflicting evidence: • External, reproducible benchmarks across a wide set of agentic tasks and non‑NVIDIA hardware are not provided in the supplied excerpts; NVIDIA's throughput and partner improvement claims come from NVIDIA (marketing/partner reporting) and are supported by limited independent test mentions (EXO Labs) but not broader third‑party benchmarking data in the excerpts. • Scope and conditions for the reported "up to 5×" throughput (which models were compared, datasets, GPU types, batch/context sizes) are not fully specified in the excerpts. • The Decoder notes that NVIDIA optimizes for its own hardware (implying potential hardware dependence), but the degree to which performance generalizes off NVIDIA platforms is not specified. • Availability timing and technical details for the Bolt 4GB variant and exact licensing/usage restrictions for released weights/datasets are not specified in the excerpts.
WHAT TO WATCH
Concrete observable signals to watch next: • Publication of the Nemotron 3.5 Super VL technical report and independent third‑party benchmark results (Artificial Analysis leaderboard entries, EXO Labs or other independent tests). • Availability of model weights, datasets, and the GitHub usage cookbook on Hugging Face/GitHub and any license terms. • Release and distribution details (dates, system requirements) for the Bolt 24GB and Bolt E4B 4GB variants. • Third‑party reports from Scale AI, Ramp, AI21 or other adopters detailing reproducible cost, latency, and accuracy metrics in production. • Tooling and deployment support for NVFP4 quantization and LiLiCorr on non‑NVIDIA inference stacks and measured performance there.
WHY IT MATTERS
A single open model that combines vision, language, reasoning, and tool use—while improving inference efficiency—can simplify enterprise agent stacks and materially reduce costs and latency for specialized workflows.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
NVIDIA announced Nemotron 3.5 Super VL as an open multimodal model with a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token for agentic workflows.
SUPPORTEDChecked 2026-10-15 · 2 supporting
NVIDIA's developer blog describes the model as 120B MoE and states 12B parameters are active per token.
SUPPORTS The DecoderMEDIA · 2026-10-15Media report restates NVIDIA's description of the 120B MoE model and the 12B active-parameter claim; this repeats NVIDIA's announcement and is not independent confirmation.
Nematron 3.5 Super VL combines reasoning, instruction following, tool use, and text, image, document, and video understanding in a single model, allowing teams to handle text and visual tasks without deploying separate vision-language models.
SUPPORTEDChecked 2026-10-15 · 2 supporting
NVIDIA states the model combines those capabilities into a single model and says this can eliminate the need for separate vision-language models.
SUPPORTS The DecoderMEDIA · 2026-10-15Media article describes the model as built to process text, images, documents, and videos for continuous enterprise agent workloads, echoing NVIDIA's claim.
NVIDIA reports that Nemotron 3.5 Super VL delivers up to 5× higher single-GPU throughput for agentic tasks compared to other open models in its size category.
SUPPORTEDChecked 2026-10-15 · 2 supporting
NVIDIA's blog states 'up to 5x higher throughput' on a single GPU for agentic tasks versus other open models in its size class.
SUPPORTS The DecoderMEDIA · 2026-10-15Media coverage repeats NVIDIA's claim of up to five times the throughput; this is a restatement of NVIDIA's announcement.
NVIDIA provides resources for post-training and customization via NeMo (supervised fine-tuning, LoRA, reinforcement learning) and publishes datasets and training recipes on Hugging Face and GitHub.
SUPPORTEDChecked 2026-10-15 · 2 supporting
The NVIDIA blog states developers can customize and post-train the model with NeMo using SFT, LoRA, and RL, and references datasets on Hugging Face and a GitHub usage cookbook.
SUPPORTS The DecoderMEDIA · 2026-10-15Media article notes NVIDIA is publishing datasets and training recipes alongside the model weights, echoing NVIDIA's disclosure.
SOURCES & TIMELINE
2NVIDIA Nemotron 3.5 Super VL is an open, customizable multimodal model with a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token for agentic workflows. The model combines reasoning, instruction following, tool use, and text, image, document, and video understanding in a single model, eliminating the need for separate vision-language models. On a single GPU, Nemotron 3.5 Super VL …
Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters built for agent tasks and document analysis in enterprise settings. Two smaller Bolt variants target local machines: a heavily compressed model for systems with 24 GB of memory or more, and a later 4 GB version for voice assistants. Nvidia is publishing datasets and training recipes alongside the model weights, but optim…