NVIDIA open-sources Nemotron 3.5 Bolt E4B, a 4GB model topping MultiChallenge

On October 23, NVIDIA released an open-source small model, Nemotron 3.5 Bolt E4B, on Hugging Face—a conversational agent that can run locally with just 4GB of memory. The model is built on Google Gemma 4 E4B (per a repost, 8B parameters with 4B effective parameters, capable of running locally on edge devices like 6GB-memory phones) via Nemotron post-training. It preserves Gemma's conversational ability while strengthening multi-turn tool calling, instruction following, and coding, and adds vision support, targeting local voice/device assistants, interactive characters, and robotics.

Confirmed

The model continues a fully open policy: weights, data, and training recipes are all freely available and customizable.
The official performance chart shows Bolt E4B in the "fast and smart" zone on a scatter plot of representative benchmark average score vs. time to first token (TTFT), scoring 70.2 with roughly 320ms TTFT, outperforming the previous generation.
NVIDIA cited ScaleAILabs statistics showing the model tops the MultiChallenge leaderboard at 56.03%, ahead of Gemma 4 E4B (the official post mentioned the latter sits around the 43% tier).

Why it matters

In the on-device small-model race, NVIDIA leveraged the Gemma base to quickly ship an open model optimized for agent scenarios, differentiating through low latency, low memory footprint, and a fully open recipe (data + recipe, not just weights), directly targeting local agent deployment on phones and other edge devices.

2026-10-23 ~ 2026-10-24 · 6 related posts

Primary sources

NVIDIA releases Nemotron 3.5 Bolt E4B, a 4GB open model for on-device conversational agents — NVIDIAAI · 2026-10-23 23:41 (UTC+08:00)
NVIDIA: Bolt E4B tops MultiChallenge at 56.03% with weights, data, recipes open — NVIDIAAI · 2026-10-24 00:05 (UTC+08:00)
NVIDIA releases open Nemotron 3.5 Bolt E4B 4GB with weights, data and recipes — NVIDIAAI · 2026-10-23 23:41 (UTC+08:00)

[source] NVIDIA releases Nemotron 3.5 Bolt E4B, a 4GB open model for on-device conversational agents — NVIDIAAI · 2026-10-23
NVIDIA's chart: Nemotron Bolt E4B tops small models at ~320ms first-token latency — NVIDIAAI · 2026-10-23
[source] NVIDIA releases open Nemotron 3.5 Bolt E4B 4GB with weights, data and recipes — NVIDIAAI · 2026-10-23
NVIDIA ships Nemotron 3.5 Bolt E4B, an open vision model that runs on 6GB RAM — danielhanchen · 2026-10-23
[source] NVIDIA: Bolt E4B tops MultiChallenge at 56.03% with weights, data, recipes open — NVIDIAAI · 2026-10-24
NVIDIA ships Nemotron-3.5-Bolt-E4B, a 4GB Gemma-4-derived SLM for low-latency on-device voice assistants — jacek2023 · 2026-10-24