Nvidia’s 4GB Nemotron Bolt model is for local assistants, game characters and robots

Nvidia has made Nemotron 3.5 Bolt E4B 4GB available, according to a post from its @NVIDIAAI account on October 23. The company calls it its newest small open model, sized for devices with about 4GB of memory.

Nvidia’s post describes the target use: “Our newest small open model is built for local assistants, interactive characters and robots that can carry on a conversation and use tools to get things done.” The post also points to a video demo. The version of the post available to us was cut off and did not include a link to the model files.

A small model aimed at conversation and tool use

The name carries the key spec. “4GB” refers to the memory footprint Nvidia is designing for, and “E4B” is the model’s size tag. Nvidia did not give a parameter count in the post. The hardware and use cases come from Nvidia’s October 15 blog post on the Nemotron 3.5 family, which listed the E4B model as coming on October 23. That post says it is optimized for “intelligence per GB” and aimed at physical and conversational agents such as voice assistants, interactive characters and embedded interfaces. It also names multi-turn agentic abilities as a goal.

The same blog post lists GeForce RTX GPUs, Jetson modules and RTX or DGX Spark systems as target hardware. In practice, that covers gaming PCs, robots and edge devices. A model of this size could run on the device itself, with no cloud call for each reply. For a character in a game or a robot, that means lower latency and no dependence on a connection.

The larger Bolt model sets the context

The E4B is the smaller of two Bolt models. According to Nvidia’s blog, the other is Nemotron 3.5 Bolt 74B-A9B 24GB. It has 74 billion total parameters, of which 9 billion are active, and uses native 2-bit quantization to fit in 24GB of memory. Nvidia says it is available now on Hugging Face, works with llama.cpp, Ollama, LM Studio, Unsloth, Canonical and vLLM, and ranks first among open models in its class on PinchBench and second on BrowseComp. Those are Nvidia’s own claims and were not independently checked here.

The pairing shows a clear split. The 74B model targets laptops and workstations running agentic tasks next to other applications. The 4GB model targets hardware with far less room, where a general chatbot would not fit.

What the announcement does not establish

Several details are missing from what Nvidia has published so far. The post gives no parameter count, context length, supported languages, license terms or benchmark results for the E4B model. The October 15 blog post offered no benchmarks for it either. The licensing question matters: Nvidia calls the model “open,” but the sources we read do not say whether that means open weights under a permissive license, or something more restricted. The blog states open weights, datasets and training recipes only for the Super VL model in the family, not explicitly for Bolt.

Availability also needs checking. The data aggregator DataLearner reported that, as of October 19, Nvidia’s Hugging Face organization had no repository for the model. It also said parameter count, context length and license were unstated at that point. Nvidia’s post says the model is available now, but developers should confirm the official model card on Hugging Face, which would carry the specs and license, before building on it.

Nvidia also gave no figures for how the model performs at tool use or conversation, the two skills it emphasizes. Claims about “robots that can carry on a conversation” come from the company’s description and a demo video. They are not independent test results.

What to watch next

Three signals will show whether the model is useful. First, the model card: parameter count, license and context length. Second, third-party benchmarks and community quantizations, which followed the 74B model quickly through Unsloth and similar projects. Third, adoption in game character tools and Jetson-based robot projects, the two uses Nvidia names. For readers following Nvidia’s wider hardware push, our report on Supermicro’s Vera CPU rack for agent sandboxes covers the data center end of the same agent strategy. The Bolt models sit at the opposite end, on the device.