Not established from the available sources.
NVIDIA · MODEL RELEASE TRACKER
Nemotron 3.5
Nemotron 3.5 Super VL is an open, customizable 120B-parameter mixture-of-experts multimodal model from NVIDIA (announced 2026-10-15) designed for continuous agentic workflows combining text, image, document, and video understanding.CURRENT SNAPSHOT3/5 DIMENSIONS WITH DATA
The dimensions that change the decision.
Not established from the available sources.
Modalities and agent capabilitiesCombines reasoning, instruction following, tool use, and understanding of text, images, documents, and video in a single model (eliminating need for separate vision-language models for these agentic workflows).
Single-GPU throughput vs. comparable open modelsOn 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).
Reported post-training improvementsOrganizations that post-trained the model reportedly achieved improvements such as up to 40× lower inference cost, 10× lower latency, and 40% higher accuracy for specialized workflows.
Throughput (single GPU)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 (per NVIDIA and reported comparisons).
Comparative accuracy and benchmarksNVIDIA states Super VL was distilled from Nemotron 3 Ultra and can match Ultra's accuracy with roughly one‑fifth the parameters; in reported agent tests (Hermes, LangChain) it scored higher than Ultra and it outperforms Nemotron 3 Super on coding benchmarks SWE‑bench and Terminal‑Bench.
Availability / opennessDescribed as an open model; NVIDIA is publishing model weights, datasets, and training recipes and provides a usage cookbook on GitHub and datasets on Hugging Face.
RELEASE TIMELINE
Published, source-backed release events only.
VERIFIABLE FACTS
Every value stays attached to a source and date.
RELEASE · Announcement dateDEVELOPER CLAIM
Announced by NVIDIA on 2026-10-15 (NVIDIA blog post and press coverage).
CAPABILITIES · Architecture and parameterizationDEVELOPER CLAIM
120B-parameter mixture-of-experts (MoE) architecture that activates 12B parameters per token.
MODALITIES · Modalities and agent capabilitiesDEVELOPER CLAIM
Combines reasoning, instruction following, tool use, and understanding of text, images, documents, and video in a single model (eliminating need for separate vision-language models for these agentic workflows).
BENCHMARKS · Single-GPU throughput vs. comparable open modelsDEVELOPER CLAIM
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).
BENCHMARKS · Reported post-training improvementsDEVELOPER CLAIM
Organizations that post-trained the model reportedly achieved improvements such as up to 40× lower inference cost, 10× lower latency, and 40% higher accuracy for specialized workflows.
CAPABILITIES · Customization and inference optimizationsDEVELOPER CLAIM
Can be customized with NVIDIA NeMo via supervised fine-tuning, low-rank adaptation, and reinforcement learning; uses techniques such as LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse to optimize inference efficiency.
AVAILABILITY · Availability / opennessDEVELOPER CLAIM
Described as an open model; NVIDIA is publishing model weights, datasets, and training recipes and provides a usage cookbook on GitHub and datasets on Hugging Face.
CAPABILITIES · Architecture and parametersDEVELOPER CLAIM
120 billion-parameter mixture-of-experts architecture that activates ~12 billion parameters per token (MoE design to balance capacity and inference efficiency).
CAPABILITIES · Customization and fine‑tuningDEVELOPER CLAIM
Developers can post-train/customize the model using NVIDIA NeMo with supervised fine‑tuning, low‑rank adaptation, and reinforcement learning on open datasets released by NVIDIA.
CAPABILITIES · Inference optimizations and deploymentDEVELOPER CLAIM
Uses LiLiCorr speculative decoding, NVFP4 quantization with quantization‑aware distillation, multi‑token prediction, and context reuse to optimize inference; optimized for NVIDIA platforms (examples include DGX, NIM).
BENCHMARKS · Throughput (single GPU)DEVELOPER CLAIM
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 (per NVIDIA and reported comparisons).
BENCHMARKS · Comparative accuracy and benchmarksDEVELOPER CLAIM
NVIDIA states Super VL was distilled from Nemotron 3 Ultra and can match Ultra's accuracy with roughly one‑fifth the parameters; in reported agent tests (Hermes, LangChain) it scored higher than Ultra and it outperforms Nemotron 3 Super on coding benchmarks SWE‑bench and Terminal‑Bench.
WHAT CHANGED
Stored passport versions, without reconstructed history.
RELEASE · Announcement dateAnnounced by NVIDIA on 2026-10-15 (NVIDIA Developer blog post).→Announced by NVIDIA on 2026-10-15 (NVIDIA blog post and press coverage).
CAPABILITIES · Architecture and parameters+ 120 billion-parameter mixture-of-experts architecture that activates ~12 billion parameters per token (MoE design to balance capacity and inference efficiency).
BENCHMARKS · Comparative accuracy and benchmarks+ NVIDIA states Super VL was distilled from Nemotron 3 Ultra and can match Ultra's accuracy with roughly one‑fifth the parameters; in reported agent tests (Hermes, LangChain) it scored higher than Ultra and it outperforms Nemotron 3 Super on coding benchmarks SWE‑bench and Terminal‑Bench.
CAPABILITIES · Customization and fine‑tuning+ Developers can post-train/customize the model using NVIDIA NeMo with supervised fine‑tuning, low‑rank adaptation, and reinforcement learning on open datasets released by NVIDIA.
CAPABILITIES · Inference optimizations and deployment+ Uses LiLiCorr speculative decoding, NVFP4 quantization with quantization‑aware distillation, multi‑token prediction, and context reuse to optimize inference; optimized for NVIDIA platforms (examples include DGX, NIM).
MODALITIES · Modalities supported− Combines text, image, document, and video understanding with reasoning, instruction following, and tool use in a single multimodal model.
MODALITIES · Modalities and agent capabilities+ Combines reasoning, instruction following, tool use, and understanding of text, images, documents, and video in a single model (eliminating need for separate vision-language models for these agentic workflows).
AVAILABILITY · Availability / openness+ Described as an open model; NVIDIA is publishing model weights, datasets, and training recipes and provides a usage cookbook on GitHub and datasets on Hugging Face.
AVAILABILITY · Openness and supporting resources− Described as open and customizable; NVIDIA released open datasets and resources on Hugging Face and published a technical report and a usage cookbook on GitHub.
BENCHMARKS · Throughput (single GPU)+ 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 (per NVIDIA and reported comparisons).