NVIDIA Releases Nemotron 3.5 Bolt E4B 4GB: An Open Small Language Model for Local Voice Agents, Game Characters and Robots

NVIDIA has released Nemotron 3.5 Bolt E4B 4GB, an open small language model built for fast conversational agents: voice assistants, game characters and robot interfaces running locally on RTX, Jetson and Spark hardware. It's post-trained from Google's Gemma 4 E4B using reinforcement learning only.

The point is headroom. An agent that holds a conversation and calls tools in roughly 4GB of VRAM leaves space on the same device for speech and vision models. That's what makes a robot or a game character feel responsive without a round trip to the cloud.

We're glad to see serious effort on small, local models. Not every assistant needs a data center, and running on the device is better for latency, for cost and often for privacy. NVIDIA's demo of a Reachy Mini robot on a Jetson Orin Nano 8GB, with a separate vision model alongside Bolt, is exactly the kind of practical stack we want to see more of.

The training choice is the interesting technical story. NVIDIA skipped supervised fine-tuning entirely because, per its report, it improved tool calling but made Gemma chattier and worse at conversation. RL with length penalties kept answers from bloating. NVIDIA also published the datasets and full recipe, which is the kind of openness that lets others check the work.

On the numbers, read carefully. Bolt led 15 of 21 benchmarks against Gemma 4 E4B, Qwen3.5-4B and Nemotron 3 Nano 4B, but all of those scores are NVIDIA-measured in NVIDIA's own harness, and the source notes vendor-reported scores differ. The standout is instruction hierarchy, IHEval 84.8 against 67.0 for Gemma. The weak spot matters for agents: on Tau2-Bench, Bolt scores 55.1 to Qwen's 75.1, and Qwen also leads on long-context tests. For a model pitched at multi-turn tool use, that gap is worth testing before you ship.

And the name. The 4GB label is a VRAM target, not the download; the quantized file on Hugging Face is 6.18 GB. On RTX, NVIDIA measured 3.1 to 4.3 GB of VRAM, so the claim is in range there, but the Jetson figures run at 4.2 GB with first-token times of several seconds on the Orin Nano. Snappy on a desktop GPU, slower on the small board.

The bigger picture: local agents in games, homes and factories will talk to people and trigger actions with little oversight. Strong instruction-hierarchy scores are encouraging there. Weaker agentic task scores are a reminder that a confident voice on a robot can still pick the wrong tool. If these spread into toys and devices, we want clear limits on what the tools can actually do.

Bolt has 8B total and 4B effective parameters with a 64K context window, a conversational mode flag for short speech-ready replies, and a configurable thinking budget to cap latency. It ships with a DFlash drafter for speculative decoding; on an RTX 5070, NVIDIA reports 272.6 tokens per second versus 225.9 for Gemma 4 E4B with its own drafter. Weights, data and recipes are released under OpenMDW-1.1, and the model is on Hugging Face, Ollama, LM Studio, Unsloth, Jetson AI Lab and as an NVIDIA NIM.