JOURNAL / 2026.10.19

Nemotron 3.5 Bolt brings agents to local machines, with conditions on its 24 GB target

NVIDIA compresses Super VL into a local model with downloadable weights. The advance lies in designing capability for limited memory; the technical report clarifies what that budget includes and what the speed tests establish.

NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on October 15. The former combines reasoning, vision and tools for specialized agents; the latter brings some of that capability to local machines through compression and quantization. Both have downloadable weights. Bolt E4B, the smaller variant announced alongside them, is scheduled for October 23: it is not yet part of the available release.

The most interesting development is how the local model is made. The Super VL model card describes 120 billion total parameters, 12 billion active per token and text, image and video inputs, with text output. Its BF16 weights provide a reference for evaluation and adaptation; a separate version is optimized for inference. Requirements for one should not be carried over to the other.

The Bolt technical explanation describes a deeper transformation than storing the same weights with fewer bits. First, it reduces Super VL's architecture to 74 billion parameters, around 9 billion active, and recovers capabilities through further training. Then it quantizes the experts to roughly two bits and trains again to compensate for the loss of precision. Other components retain higher precision. The practical idea is to start with a large model and adapt it to a memory budget, rather than treating compression as the final step before distribution.

This does not preserve every capability of the original intact. The published evaluations show trade-offs between Super VL, BF16 Bolt and quantized Bolt. I find this a useful direction precisely because those steps can be examined separately: an application might accept a small loss in general reasoning if its specific operations remain reliable, but should not assume that they do.

What actually fits within 24 GB

The Bolt report, section 8.2 estimates 23.83 GB for a configuration with one sequence, 131,072 context tokens and an auxiliary speculative decoding model. The calculation includes weights, caches and activations, but explicitly excludes the vision encoder and its projector: this target covers text serving. It is a deployment estimate, not our measurement or a guarantee for any 24 GB card.

The Bolt model card specifies up to 256,000 context tokens shared by instructions, visual inputs, history, tools and the response. It warns that recipes may reduce this according to memory and concurrency. Consequently, “multimodal model,” “maximum context” and “24 GB” do not necessarily describe the same simultaneous configuration. Working with screenshots or video also requires budgeting for the visual components.

Speed needs another careful reading. NVIDIA's performance table measures Bolt and several Qwen and Gemma models on the same 128 GB RTX Spark, using llama.cpp and their respective speculative decoding mechanisms. Bolt outperforms Qwen3.8 27B in those tests, but trails Qwen3.6 35B-A3B and Gemma 4 26B-A4B. This is no universal win. Moreover, AA-AgentPerf Local replays recorded conversations: it measures how the machine serves them, without evaluating response quality. Its timings do not establish that an agent correctly solves a fresh assignment sooner.

The weights are offered under OpenMDW 1.1. The license permits using and modifying the materials; redistribution requires retaining the agreement and relevant notices. Openness makes it easier to adapt the model and maintain a version of your own. It does not remove tool integration work or automatically make an agent private when it calls external services.

My reading is that Bolt expands a specific option: running an assistant with tools where available memory is a central constraint. To decide whether it pays off, I would compare the text configuration first, then add vision, using the same documents and operations. I would measure peak memory, success and recovery after errors. If the compressed version requires more corrections or outsourcing the difficult tasks, that cost counts too. NVIDIA's contribution deserves attention for making the path to a local agent inspectable; its usefulness will depend on the work it can still do after compression.

Primary sources consulted on October 19, 2026. We have not run these models; the cited estimates and comparisons come from NVIDIA.

← Back to journal