NVIDIA Nemotron 3.5 Bolt and Super VL: Local Agents Meet Multimodal Reasoning

NVIDIA released two related open models for agent work: Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B-24GB. Super VL is the full multimodal model for specialized agents that read text, images, documents, and video. Bolt is a compressed version of that parent, packed to run everyday agent tasks in about 24GB of memory.

Both accept text, images, and video, and reply in text. Weights, training data, and recipes ship under the OpenMDW 1.1 license. Hugging Face lists the Bolt release on October 15, 2026.

Super VL is a mixture-of-experts model with 120 billion total parameters and 12 billion active per token. The language backbone mixes Mamba-2, MoE, and attention layers with a vision encoder, so one model can handle charts, tables, page layouts, and short video clips alongside tool use and coding.

NVIDIA’s developer blog positions it for workflows such as web research, document intelligence, cybersecurity triage, and video review. Context length can reach 1 million tokens, though a given deploy often uses a shorter window to leave room for images, video frames, and concurrent requests.

You can download full-precision BF16 weights for customization and evaluation, or an NVFP4 build for leaner serving. NVIDIA’s recipes put BF16 on multi-GPU setups such as 4× H100 or 2× GB200, and list NVFP4 on 1× DGX Spark, 1× H200, 2× H100, or 1× GB200.

Bolt starts from Super VL and compresses it with Star Elastic, NVIDIA’s architecture-search and compression pipeline. The result is 74 billion total parameters with 9 billion active, stored mostly as 2-bit expert weights (IQ2) so the packed model targets a 24GB footprint.

That size is aimed at local agents: tool calling, browsing, terminal work, and everyday coding help without sending every step to a cloud model. Context goes up to 256K tokens. NVIDIA’s consumer blog ties the launch to RTX Spark and names launch-day paths through llama.cpp, vLLM, Ollama, Unsloth, LM Studio, Hermes, and OpenClaw.

On NVIDIA’s own comparison table, Bolt leads several agent and reasoning suites at that memory class, including Terminal-Bench 2.1, GPQA Diamond, BrowseComp, and WebArena (vision).

Bolt’s product name ends in 24GB for a reason: NVIDIA designed the pack to fit that budget and still leave headroom for other apps on platforms such as RTX Spark. Use the exact file you download (vLLM pack or GGUF) and the context length your runtime sets; longer chats and video still add working memory on top of the weights.

Super VL needs a larger machine. Start from NVIDIA’s deployment tables for the BF16 or NVFP4 checkpoint you chose. A unified-memory system such as NVIDIA DGX Spark (128GB coherent pool) matches NVIDIA’s single-GPU NVFP4 recipe. Discrete workstations with multi-GPU layouts follow the H100 / GB200 guidance on the model cards.

Community GGUF guides list separate planning numbers for Super VL quants that sit well above Bolt’s 24GB pack. NVIDIA’s checkpoint pages and hardware recipes remain the baseline for sizing a workstation.

CORSAIR PRO builds configure-to-order AI workstations, from compact deskside systems to multi-GPU towers. For Bolt-class local agents, look for about 24GB of usable GPU or unified memory after the OS and your agent app. For Super VL, match the BF16 or NVFP4 recipe you intend to run, including DGX Spark when you want NVIDIA’s single-node NVFP4 path.

Explore options in the CORSAIR PRO configurator, or talk through requirements with a CORSAIR PRO solutions architect.