NVIDIA Releases Open Nemotron 3.5 Bolt E4B 4GB: Post-Trained on Gemma 4 E4B to Run Local Voice Assistants, Game NPCs and Small Robots in About 4GB of VRAM

A small model for local agents that talk and act

NVIDIA's Nemotron team released Nemotron 3.5 Bolt E4B 4GB on October 23, pitching it as a small conversational-agent model optimized for intelligence per GB of memory. It is not a new architecture trained from scratch: it starts from the instruction-tuned Gemma 4 E4B from Google and is post-trained with reinforcement learning from verifiable rewards (RLVR) and from human feedback (RLHF) using NeMo-RL and NeMo-Gym, with a focus on instruction following, multi-turn tool calling and natural conversation. A week earlier, NVIDIA released Bolt 74B, a larger member of the same Nemotron 3.5 family aimed at 24GB devices.

NVIDIA names three target uses: local voice assistants for PCs, phones, homes and car cabins; interactive characters such as game NPCs and companions that have to keep a persona while taking actions; and human-machine interfaces for robots and industrial equipment. In one demo, a Reachy Mini desktop robot runs both the Cosmos 3 Edge vision model and Bolt E4B on an 8GB Jetson Orin Nano, with Bolt calling animation scripts and handling speech. In another, a blacksmith trader NPC confirms the player's payment, then calls tools on its own to check the recipe, gather materials, craft the item and hand it over.

Where the "4GB" comes from

Its features include multi-turn tool use, a conversational mode that keeps answers short and ready to be spoken, reasoning that can be switched on or off and capped, a 64K-token context, and 35 languages (the base model was pretrained on more than 140). The released weights are a Q4_K_M build quantized with llama.cpp and importance-matrix calibration, along with a DFlash drafter for speculative decoding that NVIDIA says delivers up to 21% higher throughput than Gemma 4 E4B's original drafter.

"4GB" refers to memory use in target workloads, not file size. In NVIDIA's measurements, on an RTX 5070 (12GB) with 3K input and 50 output tokens it uses 3.5GB of VRAM, with a 548 ms time to first token and 133.4 tokens per second. On an 8GB Jetson Orin Nano it uses 4.2GB, with about 5.8 seconds to first token and 15.5 tokens per second, which suits embedded uses where latency matters less.

Results: strong at following rules, weaker on long tasks

NVIDIA compared the model on 21 benchmarks in five groups against Nemotron 3 Nano 4B, Qwen 3.5 4B and Gemma 4 E4B. It says Bolt E4B scores about 40% higher on average than its previous model and about 10% higher than the other peers, and produces up to about 80% fewer output tokens in reasoning-enabled comparisons. The gains are concentrated in instruction hierarchy and consistency across turns: IHEval 84.8 (Gemma 4 E4B 67.0, Qwen 3.5 4B 60.6) and MultiChallenge 53.1 (Gemma 4 E4B 42.1).

The weak spots are in NVIDIA's own table too. On the multi-turn customer-service benchmark Tau2-Bench it scores 55.1 against 75.1 for Qwen 3.5 4B; on long-text understanding, LongBench v2 is 50.5 against 59.9, and 64K RULER is 87.0 against 95.1. In other words, it is better suited to short-turn conversational agents with clear rules than to long, complex task chains. All scores are NVIDIA's own, and two of its in-house benchmarks (Brevity Bench and When2Call-MT) have not been released yet.

How to get it

The weights, the list of training datasets and the full training recipe are public, and the Hugging Face model card lists the license as OpenMDW-1.1. It can be deployed through Hugging Face, Ollama, LM Studio, Unsloth, Jetson AI Lab and NVIDIA's In-Game Inferencing SDK. For developers who want a local voice assistant or game character, it offers a more rule-abiding and memory-frugal option than general-purpose 4B models; for customer-service-style multi-turn tasks, Qwen 3.5 4B remains the safer baseline.