NVIDIA releases Nemotron 3.5 Bolt E4B for local voice agents and robots
NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model designed for fast conversational agents running locally on GeForce RTX, Jetson, RTX Spark, and DGX Spark hardware. Post-trained exclusively from Google's Gemma 4 E4B using reinforcement learning rather than supervised fine-tuning, the model features 8B total parameters, 4B effective parameters, and a 64K token context…
Key points
- NVIDIA launched Nemotron 3.5 Bolt E4B with 8B total and 4B effective parameters, fitting in roughly 4GB VRAM.
- Trained from Google's Gemma 4 E4B using reinforcement learning without supervised fine-tuning, featuring a 64K context window.
- NVIDIA claims it won 15 of 21 internal benchmarks against competing small models, reaching 272.6 tokens/s on an RTX 5070.
According to NVIDIA's internal benchmarks across 21 evaluations, Bolt outperformed Gemma 4 E4B, Qwen3.5-4B, and Nemotron 3 Nano 4B on 15 tests in reasoning mode. It reached 84.8 on IHEval compared to Gemma's 67.0, though it trailed Qwen3.5-4B on Tau2-Bench with 55.1 versus 75.1. Using speculative decoding with a DFlash drafter on an RTX 5070, NVIDIA reports generation speeds reaching 272.6 tokens per second.
The model is licensed under OpenMDW-1.1 and is distributed on Hugging Face, Ollama, LM Studio, Unsloth, and Jetson AI Lab, alongside NVIDIA NIM and the In-Game Inferencing SDK. Although named '4GB' to reflect its memory footprint goal, the downloadable Q4_K_M GGUF file measures 6.18 GB.
The story so far
2 episodes →- NVIDIA releases Nemotron 3.5 Bolt E4B for local voice agents and robotsthis story
NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots
MarkTechPost · 24 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Moonshot AI releases Kimi K3.1 Pro, a 2.8T open-weight model closing gap with proprietary leaders · 5 src
- OpenAI brings GPT-6 Astra and Codex to Google Cloud · 23 src
- Block releases Dessa, an AI model trained only on human speech · 2 src
- Ggml-org merges webgpu Q2_0 support into llama.cpp · 2 src
- Meta launches Muse frontier model family · 2 src
Comments
via GitHub Discussions