Release Summary
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. For developers already considering NVIDIA Nemotron, the main decision is whether a 120B-total, 12B-active model combining visual understanding and tool use fits an agent workload. It accepts text, images and video and generates text. Release announcement, model card.
This assessment focuses on what changes from the earlier Nemotron 3 Super and how to choose a checkpoint and serving route. It uses official materials; no hands-on inference or independent benchmarks were performed.
What Changed
Visual evidence joins the Super agent workflow
The earlier Super checkpoint was a text model. Super VL adds a vision encoder and connector so reasoning and tool use can incorporate document pages, images and sampled video frames. That makes it relevant to agents investigating dashboards or extracting evidence from visually structured documents. It does not generate visual media, and the training architecture has no audio encoder. Training architecture.
This is a change in the Super lane, not the first multimodal capability anywhere in Nemotron: Nano Omni already handles visual and audio inputs. A speech workflow should retain that distinction rather than treat “VL” as universal modality support.
Updated agent training changes the quality case
NVIDIA describes additional software-engineering training, specialist-teacher distillation and multimodal post-training. In its paired comparison, Super VL scores 63.44% versus Ultra's 56.33% on Hermes Bench, and 87.05% versus 84.3% on LangChain evaluations. Ultra is a larger sibling, not the previous same-size Super checkpoint. These results support task-specific evaluation, not a conclusion that Super VL wins every reasoning workload. Technical article.
Precision and decoding options shape deployment
NVIDIA supplies post-trained BF16 and NVFP4 checkpoints. The quantized release uses quantization-aware distillation and mixed precision; native multi-token prediction (MTP) is complemented by a separate LiLiCorr drafter for supported configurations. The choice is between deployment recipes with different memory and decoding behavior, rather than two interchangeable files. NVFP4 model card.
Compared With Previous Version
The same-size baseline is Nemotron 3 Super. NVIDIA's earlier introduction already documented a hybrid Mamba-Transformer MoE architecture, LatentMoE, native MTP and up to 1M context. Those are inherited characteristics, not four new features introduced by 3.5. Previous Super introduction.
| Decision | Nemotron 3 Super | Nemotron 3.5 Super VL |
|---|---|---|
| Core input | Text | Text, images and video |
| Advertised parameter scale | 120B total / 12B active | Same advertised scale, with visual components |
| Maximum context | Up to 1M tokens | Up to 1M; actual serving configuration still matters |
| Decoding | Existing native MTP | Updated MTP training and additional recipe-specific drafter options |
| Checkpoint terms | NVIDIA Nemotron Open Model License | OpenMDW-1.1 for the reviewed BF16 and NVFP4 checkpoints |
| Adoption action | Existing text-agent integration | Add visual-input processing and validate the selected serving recipe |
The previous Super card and current cards establish the differences above. This is a checkpoint change; the reviewed sources do not announce an automatic replacement of an existing hosted model alias.
BF16 or NVFP4: Check the Actual Workload
The NVFP4 card publishes matched-checkpoint results with other evaluation settings held fixed:
| Official evaluation | BF16 | NVFP4 |
|---|---|---|
| SWE-bench Pro | 51.78 | 51.30 |
| Terminal-Bench 2.1 | 60.67 | 62.78 |
| WebVoyager, offline evaluation | 77.72 | 66.31 |
The results do not show a uniform quantization effect. A small change on one coding test does not establish equivalence on browser-agent tasks. This is a more useful precision decision than assuming either “quantization loses all quality” or “all performance is preserved.” Matched evaluation table.
Access and Deployment Constraints
Match hardware, precision and input support
The BF16 documentation gives example deployments on 4× H100 or 2× GB200. The NVFP4 card includes one DGX Spark, one H200, two H100 or one GB200. These are documented configurations, not interchangeable guarantees of identical context, concurrency or video support. In particular, the NVFP4 deployment table describes the DGX Spark vLLM route as text and image; other listed vLLM routes include video. BF16 recipes, NVFP4 recipes.
The 12B active figure describes computation per token, not a 12B-weight download or memory requirement. Budget for the full checkpoint and runtime state. The prompt, visual inputs, tool results, conversation history and generated response share the context window; a 1M headline is not a separate allowance for each.
Use the model-specific serving configuration
Official vLLM examples expose an OpenAI-compatible API with the served model name nemotron-3.5-super-vl. Tool calling and reasoning also require the appropriate runtime parsers; an existing client library alone does not establish compatibility. The BF16 examples name qwen3_coder for tool calls and nemotron_v3 for reasoning. Follow the exact current recipe for the selected backend rather than transferring those settings to every engine.
NVIDIA's usage cookbook provides vLLM, SGLang, TensorRT-LLM and Dynamo paths, plus LoRA, multimodal RL and a visual investigation agent. These links support implementation after the model-selection decision; this page does not reproduce a deployment tutorial.
Separate model access from service costs
The released weights can be evaluated without purchasing a bundled chat subscription. Running them still consumes hardware and operating resources, while hosted endpoints apply provider-specific charges and limits. NVIDIA also points developers to NIM and hosted access routes. No universal Nemotron API price or free hosted quota is established here. Access announcement.
What the Open Release Does and Does Not Include
The current cards name OpenMDW-1.1. Use their linked license for the chosen checkpoint rather than carrying forward the earlier Super license by assumption.
The training documentation's component-level status is narrower than a blanket “everything is open” statement: Super 3.5 SFT/RL blends are not yet published, and some agentic-teacher and vision-reasoning components are unavailable. It also says the full pipeline has not run end to end from that repository. Customization resources are available, but reproducing the released model's entire reference training process is a different requirement. Open-source status.
Who Should Upgrade / Who Should Wait
Evaluate Super VL now when an existing text agent needs to inspect document layout, charts or visual evidence within the same model workflow, and a supported deployment configuration is available. Treat precision choice as part of the agent evaluation: the published BF16/NVFP4 table shows why coding and browser tasks may lead to different choices.
Keep the existing deployment while evaluating when text-only tasks already meet requirements, the service depends on a stable hosted alias, or visual throughput at the intended concurrency has not been established. Choose another modality path when native audio is essential. Wait for additional releases when a complete public reproduction of the original training pipeline is a prerequisite.
Sources
- October 15 release announcement
- Technical overview and attributed agent evaluations
- BF16 model card
- NVFP4 model card and paired evaluation
- Previous Super checkpoint
- Earlier Super technical introduction
- Nano Omni modality reference
- Deployment and customization cookbook
- Training recipe and openness limitations
Last verified October 18, 2026. Benchmark values are attributed to NVIDIA; the adoption recommendations are editorial judgments based on those sources.
