NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, describing it as an open-weight multimodal model for specialized AI agents. At launch, NVIDIA listed Hugging Face, OpenRouter, ModelScope community and NVIDIA NIM as access channels for the model (NVIDIA’s technical announcement; launch announcement).
What NVIDIA released
NVIDIA describes Super VL as a mixture-of-experts model that combines reasoning and tool use with text, image, document and video understanding. The release also includes base, post-trained and NVFP4 checkpoints, along with datasets, training recipes and reinforcement-learning environments for customization (NVIDIA’s technical announcement).
The public model label is 120B, with 12 billion parameters active per token, according to NVIDIA. Those figures describe the launch model; the BF16 checkpoint has its own documented count and input limits.
What the BF16 checkpoint supports
NVIDIA’s NeMo AutoModel documentation lists the BF16 checkpoint at 121 billion total parameters, or 124 billion with the MTP head. It also says this checkpoint has no audio encoder and does not accept audio input. That is a useful compatibility detail if you are choosing a checkpoint for an application that needs to process audio (NVIDIA NeMo AutoModel documentation).
What NVIDIA’s performance figures measure
NVIDIA’s performance claims refer to different workloads, so their multipliers describe different things. For agentic tasks in the Artificial Analysis Intelligence Index, NVIDIA reports more than 600 output tokens per second on a Baseten endpoint. The company separately reports a 4× faster completion result for a 10,000-task coding evaluation on B200 GPUs (NVIDIA’s technical announcement).
A third result concerns serving capacity with NVIDIA Dynamo, its distributed-inference system. In a setup using 16 B200 GPUs, a target of 50 tokens per second per user, KV-aware routing, session affinity and KV-cache offload, NVIDIA reports up to eight times as many users and 85% lower first-token latency. That result describes the specified serving configuration, rather than a standalone speed score for the model.
Bolt models have different release stages
NVIDIA listed the companion Nemotron 3.5 Bolt 74B-A9B 24GB as available at launch. The company describes it as a 2-bit-quantized model with 74 billion total parameters and 9 billion active parameters, packaged for local workloads (NVIDIA’s launch announcement).
NVIDIA scheduled Nemotron 3.5 Bolt E4B 4GB for October 23, 2026 (NVIDIA’s launch announcement).
