NVIDIA released a model that handles long documents, images, and video together and predicts upcoming content. Developers building document analysis or coding assistants can choose training or deployment weights. However, long-input capacity and generation speed need separate consideration.
1. Choose the purpose before the precision
The release is NVIDIA Nemotron 3.5 Super VL 120B-A12B. All three model cards list October 15, 2026 as the release date. Commercial and noncommercial use is governed by the OpenMDW 1.1 license.
BF16 provides weights for custom training and reference evaluation. NVFP4 is a deployment version with lower memory requirements through quantization, which reduces numerical precision.
Base-BF16 is a starting point for your own post-training. The base model card also directs users to post-trained deployment versions for chat, reasoning, and tool calling. These three releases are therefore not simply three precision levels.
2. Context length and computational scale differ
The architecture combines Mamba-2, attention, and MoE, which selects only the needed expert subnetworks for computation, with a visual input processor. It lists 120 billion total parameters, with 12 billion active for computation.
The maximum context is 1 million tokens. Tokens are units of model input and output. Questions, visual inputs, conversation history, tool results, and generated answers share one window. Filling it with documents leaves less room for answers.
This size is the model's upper limit, not a guaranteed deployment setting. The card says memory and concurrency needs may require a smaller window. One DGX Spark configuration example uses 131,072 tokens.
3. Predicting ahead does not guarantee speed
Multi-token prediction, or MTP, recursively predicts upcoming tokens. Speculative decoding using it can be understood as drafting an answer first and then verifying it.
Deployment examples using built-in prediction in vLLM and SGLang do not require downloading a separate draft model.
For DGX Spark, however, the separate draft model LiLiCorr is recommended.
No specific speedup figures are provided for speculative decoding. The card says performance varies with hardware, request length, and concurrency. For comparisons with standard generation, it advises removing only speculative settings. Keep the model, task, sampling, and other server settings fixed.
4. Budget input and output together
Inputs support text, images, and up to 2 minutes of video; output is text. Documents can be supplied as page images and text. Video examples extract 2 frames per second, up to 256 frames. For 1080p video, the guidance is 1 frame per second, up to 128 frames.
Reasoning generation is toggled with enable_thinking in the chat template. Reasoning tokens and the final answer share the max_tokens budget, so budgeting only for answer length may be insufficient.
After following the card's server startup steps, use the following original command to check whether the API responds. This checks the model list, not quality on actual tasks.
curl --fail http://localhost:8000/v1/models5. Hardware savings require task-specific quality checks
The reference version's deployment table lists 4 H100 devices, while the quantized version lists 2 devices from the same family. These are different deployment options, not a comparison guaranteeing equal throughput.
Evaluations in the quantized model card held settings, evaluation tools, and parameters fixed apart from weights. Switching from reference to quantized weights lowered Terminal-Bench 2.1 from 66.70 to 60.80 and offline DeepSWE from 24.65 to 18.2.
Under the same conditions, GPQA-Diamond rose from 89.52 to 89.90, while the AA-Omniscience hallucination rate fell from 30.0 to 27.7. Lower memory demands come with the burden of checking task-specific quality changes. This is difficult to interpret as uniform performance preservation.
Some visual evaluations include items not yet integrated into public evaluation tools. The card says these used official implementations or internal evaluation tools to be released later, indicating different paths to reproduction.
Choosing among these releases means aligning training goals, actual deployment conditions, and task-specific quality, rather than focusing only on maximum context.