NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small language model built for fast local conversational agents. The model is post-trained from Google’s Gemma 4 E4B using reinforcement learning only — NVIDIA’s NeMo-RL and NeMo-Gym frameworks — and is designed to pack conversational-agent intelligence into roughly 4GB of VRAM. With 8 billion total parameters (4 billion effective), 64K context, multi-turn tool use and configurable reasoning, Bolt E4B targets voice and device assistants, interactive game characters and embedded robotic agents on GeForce RTX, Jetson, RTX Spark and DGX Spark hardware. Weights, data and training recipes are open under the OpenMDW-1.1 license with no revenue gates.
What Bolt E4B Is Built For
NVIDIA’s launch post is explicit that this is not a leaderboard model. Bolt E4B is optimized for “speed and intelligence per GB” in agents that converse with a human in the same physical or virtual space: fast local assistants for PCs, phones and cars; in-game NPCs and companions that blend dialogue with actions; and human-machine interfaces for robotics and industrial IoT. The demo pairs a Reachy Mini robot running on a Jetson Orin Nano 8GB with NVIDIA’s Cosmos 3 Edge for visual understanding — Bolt E4B calls the robot’s animation-library scripts and powers its spoken responses.
The feature set reflects the target: a conversational mode that encourages concise, speech-ready responses; multi-turn tool use with guidance for asynchronous tool calls; configurable reasoning that can be enabled for complex requests or disabled for speed, with options to cap thinking time for latency targets; 64K context for conversation history, tool definitions and application state; and support for 35 languages.
The Benchmarks NVIDIA Claims
NVIDIA reports roughly 40% higher average scores than its prior model, Nemotron 3 Nano 4B, across conversational-agent benchmarks, with up to 80% fewer output tokens in reasoning-enabled comparisons — a claim about decoding cost and response time as much as quality. Its strongest reported result is 84.8 on IHEval, which measures instruction-hierarchy conflicts. But the company’s own table shows where the model loses: Alibaba’s Qwen 3.5 4B beats it on Tau2-Bench’s multi-turn customer tasks, 75.1 to 55.1. All figures are NVIDIA-measured, and NVIDIA itself cautions that “no individual benchmark covers every capability an application needs” and that benchmark gains must translate into better behavior in the target application.
On hardware, the numbers are concrete. On an RTX 5070 with 12GB, Bolt E4B uses 3.5GB of VRAM with a 548-millisecond time to first token at batch size 1; with a 32K-token cache it needs 4.3GB. On a Jetson Orin Nano 8GB it runs at 4.2GB VRAM. The model ships as a Q4_K_M GGUF checkpoint (6.18GB) plus a BF16 version, a DFlash speculative-decoding drafter that NVIDIA says delivers up to 21% higher throughput than Gemma 4 E4B’s original drafter, and a NIM microservice — available on Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab.
Why It Matters
Bolt E4B is NVIDIA’s bet that the next wave of agents runs locally, not in a cloud API — on the RTX GPUs the company already has in hundreds of millions of PCs and the Jetson boards in robots and edge devices. By open-sourcing the weights, data and recipes with no revenue gates, NVIDIA is also making a platform play: every developer who builds a voice assistant on Bolt E4B builds on NVIDIA hardware. The release lands as Chinese rivals push their own efficiency story — Cambricon’s record quarter came on the same pitch — and as the industry’s model race splits between ever-larger frontier systems and ever-smaller models that fit in a pocket of VRAM. Kimi’s latest shows the small-model contest is genuinely competitive; NVIDIA’s answer is to own the hardware it runs on.
Background
The Nemotron line is NVIDIA’s open-model family, post-trained with the company’s NeMo reinforcement-learning stack. Bolt E4B follows the October 15 release of the larger Super VL 120B and Bolt 74B-A9B models — this 4GB release is the small, local-agent counterpart, joining a month of compact speech-and-agent models aimed at on-device deployment. The model card, full training recipe and deployment snippets are published alongside the release.
Source: NVIDIA
0 Comments