ModelsArchitectures & capability
NVIDIA releases Nemotron 3.5 Super VL, a 120B open multimodal model for enterprise agents
NVIDIA's Nemotron 3.5 Super VL is a 120B-parameter mixture-of-experts model that activates 12B parameters per token. It handles text, images, documents and video, ships with open weights, data and recipes, and is pitched as a fine-tunable base for specialized agents.
From the NVIDIA Generative AI article of October 15, 2026 and cited sources: The model mixes Mamba-2 and Transformer layers and has 512 routed experts per layer. [1] [4] [6]
SOURCES[1] NVIDIA Developer: Nemotron 3.5 Super VL developer.nvidia.com[4] NVIDIA Developer: Nemotron 3 Ultra developer.nvidia.com[6] Hugging Face model card huggingface.co
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The model mixes Mamba-2 and Transformer layers and has 512 routed experts per layer. [1] [4] [6]
- Original source
- NVIDIA Generative AI article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 15, 2026 article published
- Covered
- Oct 16 edition
- Evidence
- 10 sources across 8 sites
- 1NVIDIA Developer: Nemotron 3.5 Super VLdeveloper.nvidia.com
- 2The Decoderthe-decoder.com
- 3Artificial Analysis: Nemotron 3 Superartificialanalysis.ai
- 4NVIDIA Developer: Nemotron 3 Ultradeveloper.nvidia.com
- 5Artificial Analysis: Nemotron 3.5 Lightningartificialanalysis.ai
- 6Hugging Face model cardhuggingface.co
QuietHow it’s measured
Hacker News · 3 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The model mixes Mamba-2 and Transformer layers and has 512 routed experts per layer. It was distilled from the much larger Nemotron 3 Ultra and adds a C-RADIO vision encoder. Weights are released under the OpenMDW-1.1 license, which allows commercial use, including a 4-bit NVFP4 checkpoint that runs on one H200, one GB200 or one DGX Spark. [1] [4] [6]
NVIDIA says the model beats Nemotron 3 Ultra, which is about 4.6 times its size, on agent tests: 63.44% vs 56.33% on Hermes Bench and 87.05% vs 84.3% on LangChain evals. These are vendor-reported numbers. The headline speed of over 600 tokens per second was measured on four B200 GPUs and covers generation only, not full task time. [1] [6] [8]
On a general leaderboard, the model trails closed frontier systems by a wide margin, scoring 26 against 58 for Claude Opus 5.5. Reported wins over Claude Opus 5 and GPT-6.1 Sol each come from a separate fine-tune on its own narrow task. [2] [8]
For businesses, one model that handles both text and vision can replace a separate vision-language model feeding a language model. A free, rate-limited development API is available, and managed production through NVIDIA AI Enterprise costs $4,500 per GPU per year. Top performance depends on Blackwell GPUs, NVFP4 and NVIDIA's serving software. [1] [8]
The evidence shows an open multimodal model that runs on a single GPU, with public data and training recipes.
Read the full assessment
The likely implication is that teams could fine-tune it for narrow document, vision or agent workflows instead of paying frontier API prices, though savings remain unproven.
Executive brief
NVIDIA says its new open model, Nemotron 3.5 Super VL, beats its own Nemotron 3 Ultra on agent benchmarks, although Ultra is about 4.6 times its size (NVIDIA Developer). Super VL is a 120B-parameter mixture-of-experts (MoE) model that uses 12B parameters per token. It handles text, images, documents and video, and it ships with open weights, datasets and training recipes. NVIDIA's headline numbers are 5x the throughput of similar open models, plus partner fine-tunes that cut inference costs by 40x. These are claims by NVIDIA and its partners. On a general leaderboard, The Decoder reports that the model sits far behind closed frontier models: 26 points against 58 for Claude Opus 5.5 (The Decoder).
What changed and event timeline
Nemotron 3 Super launches
This is the text-only 120B/12B-active predecessor. Artificial Analysis gave it 36 on its Intelligence Index, ahead of gpt-oss-120b and behind Qwen3.5 122B ().
Nemotron 3 Ultra released
A 550B MoE with 55B active parameters, trained with multi-teacher on-policy distillation. It later became the teacher model for Super VL ().
Nemotron 3.5 Lightning ships
The small model in the 3.5 family scored 24 on the Intelligence Index, 9 points above Nemotron 3 Nano ().
Super VL goes live
Weights were posted to Hugging Face under the OpenMDW-1.1 license, alongside NIM, OpenRouter and ModelScope access (;).
Bolt E4B scheduled
This 4GB on-device model follows Bolt 74B-A9B, which shipped alongside Super VL (;).
Capabilities and access
- Model: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B. The release includes base, post-trained and NVFP4 (4-bit) checkpoints (NVIDIA Developer).
- Inputs: text, JPEG/PNG images, and MP4 video up to 2 minutes. Context up to 1M tokens (Hugging Face).
- Hardware: runs on 1× DGX Spark, 1× H200, 2× H100 or 1× GB200 (same source).
Read the full section
- Model: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B. The release includes base, post-trained and NVFP4 (4-bit) checkpoints (NVIDIA Developer).
- Inputs: text, JPEG/PNG images, and MP4 video up to 2 minutes. Context up to 1M tokens (Hugging Face).
- Hardware: runs on 1× DGX Spark, 1× H200, 2× H100 or 1× GB200 (same source).
- Cost: a free, rate-limited development API is available. Production through NVIDIA AI Enterprise costs $4,500 per GPU per year (VentureBeat).
- Local runtimes: LM Studio has a listing (LM Studio).
Technical analysis for researchers and developers
- Architecture: a hybrid of Mamba-2 and Transformer layers. Each layer has 512 routed experts. Images go through a C-RADIO vision encoder.
- Training:
- Distilled from Ultra using a "SkewReverse" loss.
Read the full section
- Architecture: a hybrid of Mamba-2 and Transformer layers. Each layer has 512 routed experts. Images go through a C-RADIO vision encoder. The vision encoder and embeddings stay in BF16, while the rest uses NVFP4 (Hugging Face).
- Training:
- Distilled from Ultra using a "SkewReverse" loss.
- Multi-teacher on-policy distillation: the model generates answers and specialist teachers score them on video, vision tool use and alignment.
- A two-stage supervised fine-tuning pipeline.
- 10B Indic-language tokens added in continued pretraining (NVIDIA Developer).
- Inference:
- LiLiCorr speculative decoding scores whole drafts of candidate tokens at once. NVIDIA reports up to 2.3x faster generation on DGX Spark.
- The model has a built-in multi-token prediction drafter.
- Quantization-aware distillation allows the KV cache to be stored in NVFP4.
- Prefix caching keeps both attention caches and Mamba states.
- Portability: on non-Blackwell GPUs, the same checkpoint runs with 4-bit weights and 16-bit activations.
- Reproducibility: the technical report, a usage cookbook and the datasets are public (same source).
Claims and evidence
- Throughput: NVIDIA cites "up to 5x" and over 600 tokens/s on Baseten (NVIDIA Developer).
- Benchmarks (NVIDIA-reported):
- Hermes Bench 63.44% vs 56.33% for Ultra; LangChain evals 87.05% vs 84.3%.
Read the full section
- Throughput: NVIDIA cites "up to 5x" and over 600 tokens/s on Baseten (NVIDIA Developer). VentureBeat notes the roughly 625 tokens/s figure was measured on four B200 GPUs and covers generation only, not end-to-end task time (VentureBeat).
- Benchmarks (NVIDIA-reported):
- Hermes Bench 63.44% vs 56.33% for Ultra; LangChain evals 87.05% vs 84.3%.
- CharXiv chart reasoning 80.8%.
- SWE-bench Verified 71.07 and GPQA-Diamond 89.90 (Hugging Face).
- Partner results: Scale AI reports 40x lower cost for Skydance's vendor-credit processing, and AI21/Alice report 41% higher accuracy on threat analysis.
- Beating frontier models: VentureBeat reports wins over Claude Opus 5 and GPT-6.1 Sol. Each win comes from a different custom fine-tune on a different narrow task, and NVIDIA published no task definitions or sample sizes.
- EXO Labs: NVIDIA cites an "independent" EXO Labs evaluation that puts the model on the accuracy/speed Pareto frontier on DGX Spark. The reviewed sources contain no separately published EXO report.
Context and prior work
Super VL has the same parameter footprint as Nemotron 3 Super. It adds vision and replaces separate vision-language models such as Nemotron 3 Nano Omni, which scored 63% on CharXiv (NVIDIA Developer). Baseten reports a GDPval Elo of 1190, up from 493 for Nemotron 3 Super (Baseten). The Indic dataset, about 2.8T tokens across 22 languages, was built with IIT Madras and its AI4Bharat lab.
Limitations, safety and contested findings
- General capability: The Decoder places the model well behind leading closed models, at 26 vs 58 for Claude Opus 5.5 (The Decoder).
- Fine-tune comparisons: the partner and frontier-model wins are not general claims. Each is a separate fine-tune on its own task with undisclosed methodology (VentureBeat).
- Context: the 1M-token limit may be lower in practice, depending on GPU memory (Hugging Face).
Read the full section
- General capability: The Decoder places the model well behind leading closed models, at 26 vs 58 for Claude Opus 5.5 (The Decoder). Nemotron 3 Super's 36 came from an earlier version of the index, so the two scores are not directly comparable.
- Fine-tune comparisons: the partner and frontier-model wins are not general claims. Each is a separate fine-tune on its own task with undisclosed methodology (VentureBeat).
- Context: the 1M-token limit may be lower in practice, depending on GPU memory (Hugging Face).
- Vendor incentive: The Decoder notes the models are tuned to sell NVIDIA hardware.
Business and practitioner implications
- Fine-tuning economics: the pitch is that a mid-size open model, fine-tuned on narrow workflows such as receipts, claims or threat triage, can beat frontier APIs on cost and latency.
- Simpler stacks: one model for text and vision can replace a separate VLM-plus-LLM pipeline.
- Licensing: OpenMDW-1.1 permits commercial use. Managed production support costs $4,500 per GPU per year (VentureBeat).
Read the full section
- Fine-tuning economics: the pitch is that a mid-size open model, fine-tuned on narrow workflows such as receipts, claims or threat triage, can beat frontier APIs on cost and latency. Test that on your own workloads before committing.
- Simpler stacks: one model for text and vision can replace a separate VLM-plus-LLM pipeline.
- Licensing: OpenMDW-1.1 permits commercial use. Managed production support costs $4,500 per GPU per year (VentureBeat).
- Hardware: the best performance depends on Blackwell features, NVFP4 and NVIDIA's Dynamo and NIM serving tools.