Skip to content
Neural Daily

ModelsNews

Nemotron 3.5 Bolt E4B: NVIDIA's 4 GB agent model

🇭🇺 Olvasd el magyarul →

NVIDIA released Nemotron 3.5 Bolt E4B, an open small model built on Gemma 4 E4B for local voice assistants, game characters and robots. Quantized, it runs in about 4 GB of VRAM.

By: Mikola Máté Patrik · · 2 min read

Reachy Mini robot on a table running NVIDIA Nemotron 3.5 Bolt E4B on a Jetson Orin Nano, with the robot's camera view in the corner
Image: NVIDIA · Press kit · Frame from NVIDIA's official Bolt E4B demo animation · Source

What launched?

On October 23, NVIDIA released Nemotron 3.5 Bolt E4B 4GB (NVIDIA), a further post-trained version of Google's Gemma 4 E4B using Nemotron recipes. It is not aimed at coding or leaderboards but at fast, local conversational agents: voice assistants, in-game NPCs, robots and industrial interfaces.

  • Size: 8 billion total, 4 billion effective parameters. The Q4_K_M GGUF uses roughly 3-5 GB of VRAM on target workloads.
  • 64K-token context and 35 supported languages. Reasoning can be toggled and the thinking budget capped.
  • A separate conversational mode produces shorter, speech-ready replies.
  • Multi-turn tool use, with integration guidance for asynchronous tool calls.
  • License: OpenMDW-1.1, commercial use allowed (
  • [object Object]
  • ).

What the benchmarks show

Averaged across six benchmarks (MultiChallenge, RoleMRC, BFCL-v3, IHEval, SysBench, RelyToolBench), NVIDIA reports 70.2, versus 63.2 for Gemma 4 E4B, 59.1 for Qwen 3.5 4B and 48.8 for its own Nemotron 3 Nano 4B. NVIDIA says Bolt averages about 40% higher than Nano 4B and produces up to 80% fewer tokens than peers with reasoning enabled.

NVIDIA chart of representative benchmark average versus time to first token: Nemotron 3.5 Bolt E4B 4GB at 70.2 compared with Gemma 4 and Qwen 3.5 models
NVIDIA · Press kit · Official benchmark chart from the Bolt E4B announcement · Source

Source: NVIDIA. Average score versus time to first token on an RTX 5090; circle size indicates VRAM use.

It does not win everywhere. On Tau2-Bench multi-turn customer tasks it scores 55.1% against 75.1% for Qwen 3.5 4B, and Qwen also leads on the 64K RULER long-context test. Bolt is stronger on instruction hierarchy (IHEval 84.8) and on knowing when to call a tool, clarify or abstain (When2Call-MT 63.4).

How fast is it?

On an RTX 5070 it uses 3.5 GB of VRAM, with 548 ms to first token and 133 tokens/s. On a Jetson Orin Nano 8GB it generates 15.5 tokens/s, though the first token takes almost 6 seconds. It ships with a DFlash parallel drafter that NVIDIA says delivers up to 21% higher throughput than Gemma's original MTP drafter.

Why it matters

For anyone building local assistants, kiosks or robots, latency and cost of large cloud models are the main obstacles. Bolt E4B fits on a consumer GPU alongside other models, and the full training recipe is open. It is available on Hugging Face, Ollama, LM Studio and Unsloth. Check its quality in your own language on your own prompts before committing.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles