AI BriefingNVIDIA Nemotron 3.5 Bolt E4B Released for Local Agents with 4GB VRAM
·2026.10.24 15:30
Key point
NVIDIA has released Nemotron 3.5 Bolt E4B, a small model for interactive agents that operates on 4GB VRAM.
Details
- Released an 8B (effective 4B) model that operates on 4GB VRAM (measured 3.1~5.0GB) based on RLVR post-training of Gemma 4 E4B
- Outperforms Gemma 4 E4B and Qwen3.5-4B in 15 out of 21 benchmarks based on BF16, recording a score of 84.79 on IHEval
- Trained using GRPO-based RL without SFT, achieving an average score of 63.9% in reasoning mode, with output tokens reduced by 81% compared to Qwen
- Applying DFlash speculative decoding results in a 2.74x speedup on RTX PRO 6000 and a 3.56x speedup on DGX Spark
- Supports major inference frameworks such as llama.cpp, vLLM, and Ollama, and is released under the OpenMDW license
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.