Nemotron 3.5 Super VL API

NVIDIA Nemotron 3.5 Super VL is an open vision-language model from NVIDIA for reasoning, coding, tool use, document understanding and video understanding. It combines interleaved Mamba-2, Mixture-of-Experts (MoE) and attention layers with a C-RADIO vision encoder and a vision-language connector. It has 120B total parameters with 12B active, supports contexts of up to 1M tokens and uses Multi-Token Prediction (MTP) for native speculative decoding. NVIDIA states that the model is ready for commercial and non-commercial use and releases it under the OpenMDW-1.1 license.

Nemotron is NVIDIA's family of open models with open weights, open training data and open recipes, built to deliver leading efficiency and accuracy for specialized AI agents. Nemotron 3.5 Super VL was released on October 15, 2026.

Model Summary

Property	Value
Total Parameters	120B (12B active)
Architecture	Mamba-2 and Transformer hybrid Mixture of Experts with a C-RADIO vision encoder
Context Length	Up to 1M tokens
Inputs / Outputs	Text, images and video / text
Speculative Decoding	Native Multi-Token Prediction (MTP)
Reasoning Mode	Configurable: off, on (medium) and on
Recommended Sampling	Text: temperature 1.0, top_p 0.95. Vision: temperature 0.6, top_p 0.95
Languages	English, Spanish, French, German, Italian and Japanese (model card), with multilingual reasoning and translation data also covering Chinese
License	OpenMDW-1.1

Architecture and Design

• Architecture type: Mamba-2 and Transformer hybrid Mixture of Experts (MoE) with Multi-Token Prediction (MTP).
• Network architecture: Nemotron Hybrid MoE with 512 routed experts per MoE layer, a C-RADIO vision encoder and a vision-language connector.
• Parameters: 120B total, 12B active.

The model combines sparse expert computation, attention and Mamba-2 layers to support long-context processing. The vision encoder and connector turn visual inputs into representations that the language backbone processes. A shared MTP head predicts future tokens recursively for speculative decoding. The prompt, visual inputs, conversation history, tool results and generated response all share the context window.

Inputs are text, images (RGB, jpeg and png) and sampled video frames (mp4, up to 2 minutes). Documents can be supplied as page images with accompanying text. The output is text.

Training

NVIDIA describes six training stages:

1. Foundation and Multi-Token Prediction. The language backbone is initialized from Nemotron 3 Super Base. A shared MTP head is trained for 100B tokens while the backbone stays frozen, followed by 1T tokens of distillation from Nemotron 3 Ultra Base.
2. Continued pre-training and mid-training. 100B tokens of Indic continued pre-training, a 35B-token extension to 1M context, 67B tokens of software-engineering mid-training at 256K context and a final 9B-token long-context recovery stage. The software-engineering data includes whole repositories, issue and pull-request pairs, commit diffs, build and CI logs and agent trajectories, decontaminated against SWE-bench and Terminal-Bench repositories.
3. Vision-language training. A C-RADIO vision encoder and connector extend the language backbone to images, document pages and video frames.
4. Supervised fine-tuning. A two-stage pipeline jointly trains the language and vision components on vision, reasoning, code, tools, chat, safety and long-context data, using packed sequences of up to 524,288 tokens.
5. Reinforcement learning and distillation. Reinforcement learning with verifiable rewards, reinforcement learning from human feedback and specialized teacher training improve reasoning, agentic, chat and vision capabilities. Multi-teacher on-policy distillation consolidates them into a single model.
6. MTP boosting. The MTP head is trained further against the frozen post-trained backbone to improve speculative-decoding acceptance.

Training data. The foundation corpus holds more than 20 trillion tokens (more than 10 trillion text tokens) of curated and synthetic text covering English, 19 other spoken languages and 43 programming languages, drawn from webpages, dialogue, articles, code and other material across math, science, legal, finance and general knowledge. It is built on NVIDIA's Nemotron pre-training datasets, including about 9.1T tokens of filtered Common Crawl web data, 1.7T tokens of curated and synthetic code and 660B tokens of synthetic specialized-domain data. Image training data spans 1 million to 1 billion images and video data 10,000 to 1 million hours. The foundation data was collected from 2013 to December 2025. Post-training data covers reasoning, coding, tool use, instruction following, structured outputs, chat, safety and long-context tasks together with image and video understanding. Synthetic trajectories are filtered for structural correctness, repetition and task validity.

Benchmarks

Reproduced from NVIDIA's official model card. All results for Nemotron 3.5 Super VL and the baselines were collected by NVIDIA with NeMo Gym, whose published recipes specify the containers, prompts, inference parameters, parsers and scoring for each benchmark. BabyVision, CharXiv-RQ, MathVision, MMLongBench-Doc, OCR-Reasoning, OCRBench-v2, V* and Video-MME-v2 were run with their official implementations or an internal scaffold. "n/a" means NVIDIA reports no score.

Benchmark	Nemotron 3.5 Super VL (BF16)	Qwen3.5-122B-A10B	Qwen3.8-Flash-Next	DeepSeek-V4-Flash-0731	Nemotron 3 Ultra	Inkling-Small	GLM-5.3-Flash
GPQA-Diamond	89.52	86.7	92.7	89.6	87.0	88.6	91.2
HLE, text-only	31.79	24.2	39.6	35.8	29.2	33.3	36.47
MMLU-ProX	76.91	83.2	83.0	85.5	83.1	84.0	83.0
AA-Omniscience accuracy	26.20	26.2	27.9	45.0	25.3	35.4	34.2
AA-Omniscience hallucination rate (lower is better)	30.0	84.5	41.3	95.9	22.1	59.6	30.3
AA-Omniscience index	4.03	-36.1	-1.9	-7.8	8.8	-3.1	14.3
CritPt	10.86	1.1	n/a	12.6	3.1	9.1	n/a
APEX shortlist	88.36	54.7	91.8	88.7	75.0	71.5	77.1
LiveCodeBench v6	88.22	76.7	94.1	94.6	89.2	88.3	80.6
SciCode, subtask	47.13	39.8	46.6	49.3	42.0	48.3	50.05
AA-LCR v1.1	68.62	75.1	78.4	79.7	75.2	77.3	79.5
GraphWalks	89.71	86.1	76.1	65.4	63.3	55.7	88.4
IFBench	70.86	67.6	75.6	76.2	82.9	79.3	57.7
LMArena v3 proxy	52.67	47.0	35.2	49.8	34.3	53.6	61.0
WMT24++	84.80	86.3	86.2	85.5	82.7	84.6	86.1
tau2-bench retail	69.24	55.4	71.6	73.8	73.7	71.8	78.2
tau2-bench airline	76.50	63.5	80.8	83.5	79.8	62.7	81.8
tau2-bench telecom	73.19	87.2	90.4	98.2	83.0	90.1	98.5
tau3-bench banking	18.88	14.2	44.5	38.6	18.4	n/a	34.8
GDPval-AA v2	1211	861	1561	1382	1148	1313	1563
BrowseComp	78	58.1	89.0	84.2	67.85	74.1	86.6
SWE-bench Verified	71.10	71.9	78.7	75.4	73.9	75.27	79.10
SWE-bench Multilingual	67.72	62.0	77.10	72.0	68.8	71.4	78.30
SWE-bench Pro	51.78	44.9	60.37	56.91	49.66	56.54	59.70
Terminal-Bench 2.1	60.67	48.70	85.80	76.80	54.80	53.40	82.70
BabyVision	29.60	35.7	46.8	n/a	n/a	24.7	39.1
MMLongBench-Doc	62.82	63.7	67.9	n/a	n/a	49.06	65.8
V*	85.60	91.1	93.0	n/a	n/a	70.5	88.1
OCR-Reasoning	65.08	64.3	67.1	n/a	n/a	54.7	64.6
MathVision	78.16	83.3	87.5	n/a	n/a	78.8	82.5
HLE, full set	29.20	24.4	38.8	n/a	n/a	31.6	35.1
OCRBench v2, EN	68.12	68.3	67.9	n/a	n/a	54.2	61.8
OCRBench v2, ZH	59.15	67.3	68.3	n/a	n/a	52.4	56.8
CharXiv reasoning	80.82	83.5	85.8	n/a	n/a	77.8	81.9
Video-MME v2	32.85	32.0	45.1	n/a	n/a	n/a	14.8
OSWorld-Verified	51.14	51.5	55.8	n/a	n/a	n/a	14.78
BrowseComp (offline)	78	58	89	83	44 (reported)	74	86.6
WebVoyager (offline)	77.72	54.89	90.39	not applicable	not applicable	5.25	72.64
AA-Briefcase-Lite (offline)	1137 (n/a) / 1252 (n/a) / 40	551 / 591 / 22	1636 / 1324 / 76	1469 / 1277 / 80	665 (836) / 866 (874) / 47	765 / 684 / 17	1785 (1658) / 1428 (1408) / 76
AutomationBench-AA (offline)	32.44	27.25	72.06	63.86	23.28	32.83	70.17
GDP.pdf (offline)	6.6 (AA 6.6)	8.8 (AA 3.4)	17.4 (AA 15.6)	10.8 (AA 11.0)	8.4 (AA 5.0)	8.6 (AA 11.2)	14.8 (AA 15.4)
Terminal Bench 4 (offline)	1	0	20.71	3.54	0	0.5	15.15
DeepSWE (offline)	24.65	0.4	55.57	42.3	3	18	57.77 / 63.4 (reported)

AA-Briefcase-Lite cells carry three values per model, separated here by a slash, exactly as NVIDIA publishes them. Values in parentheses are as published.

Among the models in this table, Nemotron 3.5 Super VL leads on GraphWalks (89.71) and has the second-lowest AA-Omniscience hallucination rate (30.0, behind Nemotron 3 Ultra at 22.1). Qwen3.8-Flash-Next scores higher on most coding, agentic and vision rows. Against GLM-5.3-Flash, Nemotron 3.5 Super VL scores higher on LiveCodeBench v6, IFBench, OCRBench v2 EN, OCR-Reasoning, Video-MME v2 and OSWorld-Verified.

BF16 and NVFP4. NVIDIA also publishes an NVFP4 release of the model, produced by quantization-aware distillation from the BF16 weights. With all configurations, harnesses and other non-model parameters held fixed, NVIDIA reports these scores for the two releases:

Benchmark	BF16	NVFP4
GPQA-Diamond	89.52	89.90
HLE, text-only	31.79	31.44
MMLU-ProX	76.91	73.76
AA-Omniscience accuracy	26.20	25.86
AA-Omniscience hallucination rate (lower is better)	30.0	27.7
AA-Omniscience index	4.03	5.3
APEX shortlist	88.36	89.36
LiveCodeBench v6	88.22	87.41
SciCode, subtask	47.13	46.61
AA-LCR v1.1	68.62	67.69
GraphWalks	89.71	90.34
IFBench	70.86	71.50
LMArena v3 proxy	52.67	51.93
WMT24++	84.80	84.76
tau2-bench retail	69.24	69.30
tau2-bench airline	76.50	75.75
tau2-bench telecom	73.19	73.19
tau3-bench banking	18.88	17.78
GDPval-AA v2	1211	1206
BrowseComp	78	77.5
SWE-bench Verified	71.10	71.07
SWE-bench Multilingual	67.72	67.00
SWE-bench Pro	51.78	51.30
Terminal-Bench 2.1	60.67	62.78
BabyVision	29.60	29.51
MMLongBench-Doc	62.82	62.82
V*	85.60	85.95
OCR-Reasoning	65.08	65.09
MathVision	78.16	77.79
HLE, full set	29.20	29.00
OCRBench v2, EN	68.12	68.34
OCRBench v2, ZH	59.15	59.03
CharXiv reasoning	80.82	79.88
Video-MME v2	32.85	32.54
OSWorld-Verified	51.14	52.3
WebVoyager (offline)	77.72	66.31
AutomationBench-AA (offline)	32.44	32.3
GDP.pdf (offline)	6.6	6.0
Terminal Bench 4 (offline)	1	1.5
DeepSWE (offline)	24.65	18.2

Reasoning and Sampling

Nemotron 3.5 Super VL has three reasoning settings: off, on (medium) and on. NVIDIA notes that the medium setting generates fewer tokens than full reasoning but may reduce accuracy, and that reasoning tokens and the final answer share the output budget. On Run BiOS reasoning is on by default, reasoning_effort selects medium or high and reasoning_effort: "none" turns reasoning off. NVIDIA's recommended sampling is temperature 1.0 with top_p 0.95 for text and temperature 0.6 with top_p 0.95 for vision.

Ethical Considerations

NVIDIA states that trustworthy AI is a shared responsibility. Developers should work with their internal model teams to make sure the model meets the requirements of the relevant industry and use case and addresses unforeseen product misuse. NVIDIA advises against circumventing the model's safety guardrails without a substantially similar guardrail appropriate for the use case, and publishes separate bias, explainability, privacy and safety subcards for the model.

On Run BiOS

Nemotron 3.5 Super VL is available on Run BiOS as the serverless model nvidia/nemotron-3.5-super-vl, with a 262,144-token context window, up to 131,072 output tokens, image and video input, function calling, structured outputs, optional reasoning at effort medium or high and automatic prompt caching. NVIDIA's model supports contexts of up to 1M tokens. Run BiOS serves it at 262,144 tokens so that every request can fail over between capacity sources.