Nemotron 3.5 Bolt E4B: Nvidia's 4GB Open Edge Model Built on Gemma 4

Nvidia has released Nemotron 3.5 Bolt E4B, a small open model aimed at assistants that run on the device itself: voice agents, game characters and robots. The model card on Hugging Face is dated October 23, 2026. It is a post-train of Google's Gemma 4 E4B-it with 8B total and about 4B effective parameters, shipped as a Q4_K_M GGUF that Nvidia says fits in roughly 4GB of memory.

If you read our earlier coverage of the 74B-A9B Bolt model and the RTX Spark launch software, this is the small end of the same family, the one that targets phones, Jetson boards and other low-memory hardware.

TL;DR: the questions people ask first

What exactly did Nvidia change in Gemma 4?

Nvidia describes Bolt E4B as "further post-trained from Gemma-4-E4B-it." The goal is not raw intelligence. The card says the model "carries forward Gemma's strong conversational capabilities and further improves multi-turn tool use, instruction following, and input grounding," designed for "low-memory, low-TTFT operation" so an agent can answer fast and leave memory free for other parts of the application.

TTFT means time to first token, the delay before a spoken or typed reply starts. For a voice assistant or a game character, that delay matters more than a few points on a reasoning benchmark.

Two details stand out in the card:

• A conversational mode. The chat template has a conversational_mode switch that injects an instruction to answer briefly, as in a spoken conversation. Reasoning is a separate switch (enable_thinking), with a runtime thinking budget you can cap.
• A brevity benchmark. Nvidia reports a "Brevity Benchmark" where Bolt E4B scores 82.52 in BF16 against 79.99 for Gemma 4 E4B-it and 75.35 for Qwen3.5 4B. Shorter answers mean lower latency and fewer tokens in a voice loop.

The post-training data has a cutoff of May 2026. The family is described by Nvidia as open models with open weights, training data and recipes; the card lists the post-training and pretraining dataset repos it draws on, along with some private datasets that are undisclosed.

How does it compare with Gemma 4 and Qwen3.5 4B?

The model card includes a table run by Nvidia under its NeMo Gym harness. It compares Bolt E4B with Gemma 4 E4B-it, Qwen3.5 4B and Nemotron 3 Nano 4B, in BF16 and Q4_K_M. Here are selected rows for the Q4_K_M build, the one most people will run:

Reading it honestly:

• Against its own base, Bolt wins nearly everywhere. Tau2-Bench jumps from 38.72 to 52.32 and IFBench from 42.67 to 64.00 for the quantized builds. That is the point of the post-train.
• Against Qwen3.5 4B, it is a split decision. Bolt leads on instruction hierarchy (IHEval), role play, multi-turn constraints and hallucination-under-tools tests, while Qwen3.5 4B leads on Tau2-Bench, long-context RULER, LongBench v2 and general function calling.
• Long context is the weak spot. On RULER at 64K, Bolt E4B scores 85.76 against 95.37 for Qwen3.5 4B. If your assistant must reason over long documents, test before committing.
• The numbers are Nvidia's. The card notes accuracy was measured by NVIDIA and "may differ from vendors' self-reported numbers." Most rows were run in reasoning mode, except RoleBench, EQ-Bench v3 and RULER.

Nvidia's research page also publishes accuracy versus TTFT and accuracy versus verbosity charts across open models; they are on the Bolt E4B research page, which we could not mirror, so check them there.

How to run Nemotron 3.5 Bolt E4B locally

The card's quick start uses the Hugging Face CLI and llama.cpp:

export MODEL_REPO=nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-4GB
export MODEL_FILE=Nemotron-3.5-Bolt-E4B-Q4_K_M.gguf
hf download "$MODEL_REPO" "$MODEL_FILE" --local-dir ./bolt-models

Nvidia then shows a Docker command that serves the file with the llama.cpp CUDA server image, with -c 65536, -ngl 999 and --jinja so the custom chat template and tool-call format work. A second command adds the DFlash drafter with --spec-type draft-dflash and --spec-draft-n-max 7. The card marks the llama.cpp deployment recipe as "pending," so expect the instructions to change.

Settings worth copying:

• Sampling: temperature 1.0 and top_p 0.95 in both thinking modes.
• Reasoning: turn enable_thinking on for tool-heavy agent steps and off for quick chat.
• Conversational mode: use it for voice, where short spoken-style answers cut latency.
• Unsloth: the Unsloth docs also list the model and say it runs locally on about 6GB of RAM.

If speculative decoding is new to you, our explainers on Nvidia's open DFlash 2 draft model and the Dory 4B looped model cover the surrounding Nvidia research.

Where Nvidia says it should be used

The card names local voice and device assistants, interactive personas and embedded interfaces. Nvidia's demos, as relayed by the company's posts, include a Reachy Mini robot running on a Jetson Orin Nano 8GB together with the Cosmos vision model, and a test of the model as a blacksmith trader NPC in a game. The model is text-only on the card, so the vision in the robot demo comes from the paired Cosmos model, not from Bolt itself. Third-party coverage describes Bolt as supporting vision because Gemma 4 does; the card lists text input and output, so verify the modality you need before building on it.

For builders, the pattern is a small local model that handles the conversation and routes hard steps elsewhere. Nvidia has published a related approach in Edgewise, where a local small model delegates critical steps to a cloud LLM, and its Flora open models target the robot side.

What people are asking

Is this the same as the Bolt 74B model? No. The 74B-A9B model is a large mixture-of-experts release for 24GB-class GPUs, while E4B is a dense-style small model built on Gemma 4 for phones and boards. They share the Bolt name and the focus on conversational, tool-using agents, but they have different bases, sizes and memory needs.

Why post-train Gemma instead of training from scratch? Gemma 4 E4B-it already has strong conversational behavior. Nvidia's contribution is targeted post-training data and recipes for tool use, instruction hierarchy and grounding. That is cheaper than pretraining and gives you a model whose base license and behavior you can inspect.

Should I use the BF16 or the Q4_K_M build? Nvidia publishes both and shows both in the table. The quantized scores are close to BF16 on most rows and sometimes higher, such as ToolSandbox (74.13 versus 68.77), which is within what you would expect from noise and harness differences. On Tau2-Bench the quantized build drops from 55.10 to 52.32. If memory allows, test both.

Does it do vision or audio? The Hugging Face card lists text in and text out. Nvidia's voice and robot demos pair it with other models for speech and vision, so treat Bolt as the language brain and build the rest of the pipeline around it.

A green chip inside a cutaway handheld device, illustrating a small Nemotron 3.5 Bolt E4B model running on the device

Gaps and things to verify

• Memory claims differ. "4GB" is a VRAM figure for the quantized weights; the file itself is about 6.2GB and Unsloth quotes 6GB of RAM. A phone with 8GB total may be tight once the OS and KV cache are counted.
• Context length is inconsistent across sources. The card says up to 64K tokens, Unsloth also lists 64K, but the GGUF metadata on Hugging Face reports a 131,072 context length. Use 64K, as the card recommends, unless you test further.
• License nuance. OpenMDW-1.1 is permissive, but the base model has its own terms. Read both before shipping a product.
• Benchmarks are first-party. No independent replication existed on release day, and tool-use scores vary with harness and prompt format.
• Not a coding or math model. The card pitches conversation and tool use, not frontier reasoning.

What this means for what you build

If you build voice agents, NPCs or on-device assistants, Bolt E4B is a credible new default to test against Gemma 4 E4B and Qwen3.5 4B, mainly because of the instruction-hierarchy and tool-hallucination results and the short-answer behavior. If you need long-context retrieval or the strongest agentic planning at 4B, Qwen3.5 4B still looks stronger in Nvidia's own table. Run both on your own traces.

Larger siblings are covered in our Nemotron 3.5 Super VL post.

Related reading

• Nemotron 3.5 Bolt 74B-A9B: 24GB, Star Elastic and a 2-bit drafter
• RTX Spark launch software: Nemotron 3.5 Bolt, PAIR and Llama on Windows
• Nemotron 3.5 Super VL: open 120B multimodal agent model
• Nvidia Edgewise: local small model plus cloud LLM
• Nvidia open DFlash 2 draft model for speculative decoding
• Nvidia Flora open robot data models

Model details are accurate as of October 23, 2026, from the Hugging Face model card and Unsloth's documentation; repo contents and instructions may change.