Nemotron 3.5 Super VL is an open-weight vision language model released by NVIDIA on October 15, 2026. It combines text, image, document, and video understanding with reasoning and tool use for agents. NVIDIA positions it for specialized enterprise workflows, including document analysis and software engineering. It belongs to the Nemotron family; Nemotron 3.5 Lightning is a separate model.[1]

Nemotron 3.5 Bolt 74B-A9B is a separately released compressed derivative of Super VL.[18]

Specifications

Property	Release specification
Published parameter count	120 billion total; 12 billion active per token[2]
Language architecture	Hybrid Mamba-2, attention, and mixture-of-experts layers[2][3]
Routed experts per MoE layer	512[3]
Routed experts selected per token	22, with a separate shared expert[3]
Language hidden size	4,096[3]
Attention heads	32 query heads; 2 key-value heads[3]
Vocabulary size	131,072[3]
Maximum position setting	1,048,576 tokens[3]
Inputs	Text, images, and video[2]
Output	Text[2]
Audio encoder	None in the released configuration[3]
License	OpenMDW-1.1[2]

The release advertises a context window of up to one million tokens. Serving recipes use different precision, cache, and concurrency settings, so the maximum context is not a promise that every deployment can use it.[9]

Architecture and training

The language backbone interleaves Mamba-2, attention, and mixture-of-experts blocks. A C-RADIO vision encoder and connector supply visual representations to that backbone. The official usage cookbook identifies the encoder as C-RADIO v4-H.[3][5]

NVIDIA initialized Super 3.5 from Nemotron 3 Super Base. Its technical report describes a 100-billion-token warm-up for a shared multi-token prediction head with the backbone frozen, followed by one trillion tokens of distillation from Nemotron 3 Ultra Base. Subsequent stages include Indic continued pretraining, extension to a one-million-token context, software-engineering mid-training, and long-context recovery. The software-engineering mixture includes repositories, issue and pull-request pairs, commit histories, build logs, and agent trajectories.[4]

Post-training combines supervised fine-tuning, reinforcement learning with verifiable rewards, reinforcement learning from human feedback, specialized teachers, and multi-teacher on-policy distillation. The report treats Super as a multimodal model and Lightning as text-only; their capabilities and training results are not interchangeable.[4]

Release variants

The release provides separate base, post-trained, and quantized checkpoints. The LiLiCorr download is an auxiliary drafter, not a replacement for the target model.[4][7]

Checkpoint suffix	Role
`120B-A12B-Base-BF16`	Base checkpoint[4]
`120B-A12B-BF16`	Post-trained checkpoint[4]
`120B-A12B-NVFP4`	Post-trained, mixed-precision quantized checkpoint[2]
`120B-A12B-NVFP4-LiLiCorr`	Speculative drafter loaded alongside the target model[7]

The NVFP4 release uses quantization and quantization-aware distillation. Its precision is mixed rather than uniformly four-bit: the model card specifies BF16 for the vision encoder, connector, Mamba convolutions, and embeddings, and FP32 for the MoE router.[2]

Published evaluations

The following results are NVIDIA's model-card measurements, accessed October 16, 2026, not independent replications. The card says the BF16 and NVFP4 runs keep harnesses, configurations, and other non-model parameters fixed. Values are reproduced on the source's reported scales.[2]

Benchmark	BF16	NVFP4
GPQA-Diamond	89.52	89.90
HLE, text-only	31.79	31.44
LiveCodeBench v6	88.22	87.41
SciCode, subtask	47.13	46.61
SWE-bench Verified	71.10	71.07
SWE-bench Pro	51.78	51.30
GraphWalks	89.71	90.34
IFBench	70.86	71.50
CharXiv reasoning	80.82	79.88
MMLongBench-Doc	62.82	62.82
MathVision	78.16	77.79
OCRBench v2, EN	68.12	68.34
Video-MME v2	32.85	32.54

NVIDIA publishes evaluation recipes in NeMo Gym. Each recipe pins benchmark settings and a Gym commit; the default shared configuration enables thinking with temperature 1.0 and top-p 0.95 unless overridden. SWE-bench runs use OpenCode inside a sandbox. These details matter when comparing results with a different agent harness or serving configuration.[6]

The reproducibility guide also identifies results not covered by its recipes, including several visual benchmarks run with official implementations or internal scaffolding. It distinguishes those from Arena human-vote ratings and internally harnessed results. A benchmark score therefore does not establish a universal ranking across all workloads.[6]

Independent evaluation

Artificial Analysis reported an Intelligence Index score of 26.5 in its October 15, 2026 evaluation. Its release-day analysis also found about 60% fewer output tokens per Index task than Nemotron 3 Super. The NVFP4 version scored within 0.1 Index points of BF16, with no component changing by more than 2.6 points.[13]

The evaluator's v4.3.2 methodology combines ten evaluations and is primarily text-based and English-language. Image, speech, and multilingual evaluations are reported separately. The aggregate score is therefore not a standalone measure of visual understanding.[14]

Serving and inference

The official cookbook includes vLLM, SGLang, and TensorRT-LLM deployment notebooks. It also includes a visual-investigation agent that reads dashboards and scanned documents and proposes a change for a person to approve.[5]

Two speculative decoding paths are available. Native MTP uses the target model's internal drafter. The separate LiLiCorr checkpoint is intended for DGX Spark and low-concurrency data-center deployments.[7]

LiLiCorr scores candidate relationships across adjacent draft positions so a parallel drafter can propose more coherent token blocks. The target model verifies the candidates; with correct speculative verification, this preserves its output distribution. Acceptance length and latency still depend on the workload, sampling, and serving stack.[7][8]

For clustered serving, NVIDIA Dynamo publishes model-specific vLLM and TensorRT-LLM recipes with aggregated or disaggregated worker topologies. These use the nemotron_v3 reasoning parser and qwen3_coder tool-call parser. The deployment instructions require a shared model-cache volume and adaptation to the cluster's storage, scheduling, and networking.[9]

The Dynamo performance table is tied to a SWE-bench Pro trace-replay workload, not arbitrary chat traffic. Its optimization target includes more than 50 output tokens per second per user and a median time to first token below five seconds.[9]

NVIDIA separately reports that, on 16 B200 GPUs at 50 tokens per second per user, session-affinity KV-aware routing and KV-cache offload served up to eight times as many users while reducing first-token latency by 85%. This is a serving-system result under that setup, not a hardware-independent speed guarantee.[10]

Hosted prototype API

NVIDIA's Build catalog offers a hosted prototype endpoint through https://integrate.api.nvidia.com/v1, with the model identifier nvidia/nemotron-3.5-super-vl-120b-a12b. On October 16, 2026, the catalog described an early-access model with 262K context and 124B parameters. This listing differs from the released checkpoint's published 120B parameters and up-to-one-million-token context above.[15][2]

Community GGUF conversions

Unsloth publishes third-party GGUF downloads of Super VL, separate from NVIDIA's checkpoint releases. Its repository names NVIDIA's BF16 checkpoint in its base_model metadata and includes a vision-projector file.[16]

Unsloth's local-running guide gives the following memory planning estimates, accessed October 16, 2026:[17]

GGUF variant	Download size	Estimated minimum memory
UD-IQ1_M	46.1GB	64GB
UD-Q3_K_XL	65.8GB	96GB
UD-Q4_K_XL	86.9GB	128GB
Q8_0	132GB	192GB

These are estimates for modest context lengths, including runtime, operating-system, cache, and vision-projector headroom, not measured throughput results. Memory can be split across system RAM and GPU VRAM; unified memory counts once. Longer contexts and video require more memory. The guide's llama.cpp image and video examples load the matching vision projector; PDF pages are first rendered to images.[17]

Customization

The weights can be adapted with NVIDIA NeMo. The usage cookbook provides financial question-answering LoRA training over document images and text, plus a multimodal reinforcement-learning example using synthetic star-count images. These are example workflows rather than evidence of accuracy on an organization's own documents.[5]

NeMo AutoModel documents receipt-field extraction with CORD-v2, text-to-SQL adaptation, continued pretraining on raw text, and text-only supervised fine-tuning with the vision stack frozen. Its CORD-v2 LoRA recipe freezes the base weights and trains selected language-model projections; this is different from updating the full model.[11]

Licensing and limitations

OpenMDW-1.1 grants rights to use and modify the supplied model materials, subject to its terms. Redistribution must retain the agreement and applicable origin notices. The license disclaims warranties and leaves users responsible for relevant third-party rights; it does not guarantee that a particular application is legally cleared.[12]

The target model can still produce inaccurate, incomplete, biased, or otherwise undesirable responses. NVIDIA's drafter documentation calls for application-specific safeguards and evaluation in the intended deployment environment. Speculative decoding changes how candidate tokens are generated and checked; it does not independently fact-check the final answer.[7]

References

Kari Briski, NVIDIA Blog. Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models. October 15, 2026.
NVIDIA, Hugging Face. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4: model card. Released October 15, 2026; accessed October 16, 2026.
NVIDIA, Hugging Face. Released NVFP4 checkpoint configuration. Accessed October 16, 2026.
NVIDIA. Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI. Technical report dated October 14, 2026.
NVIDIA-NeMo, GitHub. Nemotron-3.5 Super VL usage cookbook. Accessed October 16, 2026.
NVIDIA-NeMo, GitHub. Reproducing the published evaluation results. Accessed October 16, 2026.
NVIDIA, Hugging Face. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr: model card. Released October 15, 2026; accessed October 16, 2026.
Matan Rusanovsky et al. LiLiCorr: Lightweight Likelihood Correlation of Parallel Drafts for Speculative Decoding. arXiv:2608.20530, submitted August 20, 2026; revised September 22, 2026.
NVIDIA Dynamo documentation. Nemotron 3.5 Super VL deployment recipes. Development-version documentation, accessed October 16, 2026.
Chris Alexiuk and Chintan Patel, NVIDIA Technical Blog. NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning. October 15, 2026.
NVIDIA NeMo AutoModel documentation. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16. Nightly documentation, accessed October 16, 2026.
OpenMDW. OpenMDW License Agreement, version 1.1. Accessed October 16, 2026.
Artificial Analysis. NVIDIA Releases Nemotron 3.5 Super and Nemotron 3.5 Bolt. October 15, 2026.
Artificial Analysis. Intelligence Benchmarking Methodology. Version 4.3.2, accessed October 16, 2026.
NVIDIA Build. Nemotron 3.5 Super VL 120B-A12B: prototype API and catalog listing. Accessed October 16, 2026.
Unsloth, Hugging Face. NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF: model repository. Accessed October 16, 2026.
Unsloth Documentation. NVIDIA Nemotron 3.5 Super VL: Run Locally Guide. Accessed October 16, 2026.
NVIDIA, Hugging Face. NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF: model card. Released October 15, 2026; accessed October 16, 2026.

Editing: Nemotron 3.5 Super VL