NVIDIA releases Nemotron 3.5 Super VL and a 24GB Bolt model for local agents

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open vision-language model with 120 billion parameters aimed at agents that run all day. The same post, by Kari Briski, adds two smaller "Bolt" models for local machines. Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB are available now, and Nemotron 3.5 Bolt E4B 4GB follows on October 23.

What is new in Nemotron 3.5 Super VL

The Super VL model card (opens in a new tab) describes a hybrid Mamba-2 and Transformer mixture-of-experts model with 120B total and 12B active parameters. Context length goes up to 1M tokens. The model takes text, images and video as input and returns text. Video clips are limited to MP4 files of up to 2 minutes.

The published checkpoint uses NVFP4 weights. A BF16 version sits in a separate repository. NVIDIA's technical blog post (opens in a new tab) says Super VL was distilled from Nemotron 3 Ultra. The announcement (opens in a new tab) gives the training scale as about 26 trillion tokens, followed by post-training on roughly 3.8 million samples and reinforcement learning.

NVIDIA reports these results against its own earlier models:

• Hermes Bench: 63.44%, against 56.33% for Nemotron 3 Ultra.
• LangChain Evals: 87.05%, against 84.3% for Nemotron 3 Ultra.
• CharXiv, a chart reasoning test: 80.8%, against 63% for Nemotron 3 Nano Omni.

The announcement also calls Super VL 5x faster than the next best open model in its class. The comparison models are not named in the post.

Nemotron 3.5 Bolt: the local models

Bolt 74B-A9B 24GB is a compressed version of Super VL. Its GGUF model card (opens in a new tab) says NVIDIA built it with a router-guided compression method it calls Star Elastic. The result has 74B total and 9B active parameters, keeps image and video input, and supports up to 256K tokens of context.

The 24GB in the name comes from heavy quantization. All experts use IQ2-XXS at 2.0625 bits per element. The card says the design targets a footprint below 24 GB at batch size 1 and 128K context. Supported GPU generations are Ada Lovelace and Blackwell, on Windows and Linux.

NVIDIA says Bolt 74B-A9B is the top open model in its class on PinchBench and second on BrowseComp, without giving scores in the post. Bolt E4B 4GB is meant for voice assistants, interactive characters and embedded interfaces on GeForce RTX, Jetson and RTX or DGX Spark hardware. The announcement gives no architecture details for Bolt E4B, which is due on October 23.

Who the Nemotron 3.5 models are for

Super VL suits teams that self-host agents reading images, PDFs, charts and webpages. The Super VL card lists one DGX Spark, one H200, two H100s or one GB200 as deployment options. Readers weighing a DGX Spark can see the 64GB DGX Spark announcement for that hardware.

Bolt 74B-A9B is for developers who want a multimodal agent model on a single workstation or laptop GPU. NVIDIA says it is built for RTX Spark laptops, which arrive October 16, and also runs on GeForce RTX PCs and RTX PRO workstations.

Both models use the OpenMDW License Agreement, version 1.1. The Super VL card states the model is ready for commercial and non-commercial use. NVIDIA also says the release includes open datasets on Hugging Face, plus training recipes.

How to try Super VL and Bolt

Super VL is on Hugging Face, OpenRouter, the ModelScope community and as an NVIDIA NIM container. NVIDIA says the NIM is refreshed every two weeks with security patches and performance work. For self-hosting, the model card names the vllm/vllm-openai:v0.32.0 Docker image and SGLang. The guide to serving an open model with vLLM covers the OpenAI-compatible endpoint that setup gives you.

Bolt 74B-A9B works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM, according to the announcement. The GGUF card gives this Ollama command:

ollama run hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

If you are new to these runners, see how to run a local LLM with Ollama and the comparison of llama.cpp, Ollama and LM Studio.

Both model cards say reasoning is switched on or off through the chat template with an enable_thinking flag.

Limits and open questions

• Every benchmark figure comes from NVIDIA. The Bolt PinchBench and BrowseComp rankings have no published scores in the announcement.
• The "5x faster" claim names no comparison models. The technical blog's own chart caption gives a 4x figure for coding task completion on B200.
• The Dynamo serving claim (up to 8x more users and 85% lower first-token latency) was measured on 16 B200 GPUs at 50 tokens per second per user. The baseline setup is not clearly stated.
• Training token counts differ by page. The announcement says about 26 trillion tokens. The model card describes a foundation corpus of over 20 trillion tokens.
• No pricing appears for NIM or OpenRouter access in the announcement.
• Bolt E4B has no release page yet beyond the October 23 date.