Nvidia releases Nemotron 3.5 Bolt E4B, an open voice-agent model it says fits in about 4 GB

Nvidia has released Nemotron 3.5 Bolt E4B 4GB, a small open-weights language model meant to run locally on GeForce RTX cards and Jetson modules as the conversation engine for voice agents, game characters and robots. The model was created by further training Gemma 4 E4B, and Nvidia says that in Q4_K_M quantization it uses between 3.1 and 4.3 GB of memory on RTX hardware.

What Nvidia released

Nemotron 3.5 Bolt E4B is a text model that Nvidia optimized primarily for conversational agents. The company describes it as tuned for speed and intelligence per gigabyte. It supports multi-turn tool calling, a conversational mode with short answers, and configurable reasoning.

The starting point was the instruction-tuned Gemma 4 E4B. Nvidia added post-training with reinforcement learning (RLVR and RLHF) using the NeMo-RL and NeMo-Gym libraries. Every sample was used in both reasoning and non-reasoning modes, at a 2:1 ratio in favor of non-reasoning, and in one environment the length penalties were raised to curb verbosity.

The release includes several files: the post-trained model in GGUF format, a DFlash drafter for that version, the full BF16 checkpoint, and a DFlash-NVFP4 drafter for the Hugging Face checkpoint. Nvidia says the DFlash drafter delivers up to 21 percent higher throughput than the original MTP drafter from Gemma 4 E4B. The training recipes are public, while the datasets are only partly open: a sample is freely available, and the remaining data (code, math, multilingual) requires approval. The Brevity Bench and When2Call-MT sets are to be released later.

Results according to Nvidia

Nvidia says the model has about 40 percent higher average quality than its earlier Nemotron 3 Nano 4B and, in comparisons with reasoning enabled, produces up to about 80 percent fewer output tokens than competitors. These are the manufacturer's own measurements, taken on BF16 weights across a benchmark set covering conversation, tool use, instruction following and long-context robustness.

The comparison with Gemma 4 E4B, Qwen 3.5 4B and Nemotron 3 Nano 4B is mixed. Bolt scored 84.8 on IHEval against 67.0 for Gemma 4 E4B, 67.0 on RoleBench against 49.9 for Gemma, and 72.1 on LongMemEval. It loses, however, on Tau2-Bench (55.1 versus 75.1 for Qwen 3.5 4B), on LongBench v2 (50.5 versus 59.9 for Qwen) and on the RULER test at 64k tokens (87.0 versus 95.1 for Qwen and 91.5 for Nano 4B).

Hardware and use cases

In Nvidia's measurements, with Q4_K_M quantization and a batch size of 1 in llama.cpp, the model used 3.5 GB of VRAM on a 12 GB RTX 5070 and generated 133.4 tokens per second; the company cautions that results may vary with hardware, software and configuration. On RTX Spark it managed 67.0 tokens per second, and on an 8 GB Jetson Orin Nano 15.5 tokens per second with a time to first token of 5785 ms. The slower edge modules clearly lag behind, which matters for projects that need fast voice responses.

Nvidia points to device and voice assistants, interactive game characters, and embedded and industrial agents as use cases. Demonstrations showed a Reachy Mini robot running on a Jetson Orin Nano 8GB, with the Cosmos 3 Edge model handling vision, and a blacksmith-trader game character that verifies payment and then calls tools to craft and hand over an item.

What it means for builders

For Polish teams building voice assistants, game bots or edge-device apps, the main points are that the model runs locally, without sending conversations to the cloud, and is available in popular tools such as Ollama and LM Studio. It is released under the OpenMDW 1.1 license, and the base Gemma 4 is covered by the Apache 2.0 license; according to the model card, it can be used commercially and non-commercially. It is worth testing quality in Polish, though, as the sources say nothing specific about it.

The results come from the manufacturer and have not been confirmed here by independent tests. The model was designed for conversation and tool calling, but in some long-context and tool tests (Tau2-Bench, LongBench v2, RULER) it performs worse than Qwen 3.5 4B.