NVIDIA’s Nemotron 3.5 Bolt E4B Puts a Free Local Agent Model in 4GB. What About Arabic?

On October 23, NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model tuned for conversational speed and what the company calls “intelligence per GB.” Built on Google’s Gemma 4 E4B with NVIDIA’s own post-training, it is aimed at voice assistants, game characters and robot interfaces that have to run on local hardware. Our take: it is a free, practical option if you are building a local assistant, but Arabic was not one of its post-training focus languages, so test it before you trust it with Arabic users.

What changed: an agent model squeezed into about 4GB

According to the Hugging Face model card, the model has 8B total parameters (4B effective), ships in Q4_K_M quantization, takes text in and gives text out, and handles up to 64K tokens of context. Its post-training data runs to May 2026. NVIDIA says it is designed to fit conversational-agent work into as little as about 4GB of memory, leaving room for other apps or models on the same machine.

Three features matter for builders:

• Multi-turn tool use: it can call your app’s APIs and carry the results through a conversation.
• Conversational mode: a toggle that pushes it toward short, speech-ready replies.
• Controllable reasoning: thinking can be switched on or off, with an optional thinking budget to keep latency down.

NVIDIA also published a full-precision BF16 version and a companion “DFlash” draft model for speculative decoding, which speeds up generation. The company claims quality gains over its previous small model and shorter answers than rivals, but those come from its own benchmarks and have not been independently checked.

Who it is for

NVIDIA’s target list is specific: voice and device assistants, interactive characters such as game NPCs and companions, and robot or IoT interfaces. Supported hardware spans GeForce RTX cards, Jetson boards, RTX Spark and DGX Spark. The Ollama library page adds that the E4B build runs on Windows, Linux and Apple Silicon. If you just want a general chatbot, larger open models will usually serve you better; Bolt E4B makes sense when memory and response time are the constraint.

What it costs

Nothing to download. The weights are free and NVIDIA lists no price. The model is released under the OpenMDW License Agreement v1.1, while the underlying Gemma 4 base is under Apache 2.0, and the model card says it is ready for commercial and non-commercial use. Your real costs are hardware and electricity, not a subscription. As always, read the licence yourself before shipping it in a product.

How to try it now, and the Ollama tag trap

MarkTechPost reports that the model is available through Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab. The quickest route for most people is Ollama:

ollama run nemotron-3.5-bolt:e4b

Do not drop the tag. Plain ollama run nemotron-3.5-bolt pulls :latest, which is a different and far bigger model, Bolt 74B-A9B: 25GB, released about a week earlier, and not supported on Apple Silicon. Note too that Ollama’s e4b package is not identical to NVIDIA’s file. Ollama lists it at 7.2 to 7.9GB with a 128K context and image input, while NVIDIA’s 4GB file is text-only with 64K. Check the size against your free memory before you download. If you are still choosing a local runner, our Ollama vs LM Studio comparison walks through the trade-offs.

The Arabic question: listed, not trained for

NVIDIA says the model supports 35 languages on top of a base pre-trained on more than 140, and the model card’s language list includes Arabic. But the same card names the primary post-training languages as English, French, German, Italian, Japanese, Spanish and Chinese. Arabic is not among them, and NVIDIA published no Arabic-specific results. Whatever Arabic ability it has is most likely inherited from Gemma, which means dialogue quality and tool calling in Arabic are open questions.

Before you build on it for Arabic speakers, spend ten minutes on this test:

• Register and dialect: ask the same question in Modern Standard Arabic and in a dialect (Gulf, Egyptian, Levantine or Maghrebi) and see whether it answers in kind or drifts into MSA or English.
• Tool calling in Arabic: give it a booking tool and a request like “book me an appointment Thursday at 4”, then check that the day and time land in the right arguments.
• Conversational mode: switch it on and check whether short Arabic replies stay grammatical.
• A baseline: put the same prompts to an Arabic-first open model such as Jais 3. The gap will tell you more than any launch post.

One practical tip: Arabic text can display in the wrong order in some terminal windows, so judge the answers after copying them into a text editor or a graphical chat app, not by how they look in the console.