NVIDIA Releases Nemotron 3.5 Bolt E4B for Local Voice Agents, Games and Robots

NVIDIA released Nemotron 3.5 Bolt E4B on October 23, 2026, an open small language model built to run locally on robots, game characters and voice assistants, according to MarkTechPost. The model, post-trained from Google's Gemma 4 E4B using reinforcement learning only, fits in as little as 3.1GB of VRAM and led 15 of 21 benchmarks against rival small models including Qwen3.5-4B.

What Happened

According to MarkTechPost, Nemotron 3.5 Bolt E4B carries 8B total parameters with 4B effective parameters and a 64K context window, and is aimed at three workloads: device and voice assistants, interactive game characters, and embedded or industrial robot agents. NVIDIA skipped supervised fine-tuning entirely, training the model instead through reinforcement learning from verified rewards and human feedback using its NeMo-RL and NeMo-Gym tools, after finding that supervised fine-tuning hurt Gemma's conversational quality even as it improved tool-calling accuracy.

The launch report includes a demo of a Reachy Mini robot running on a Jetson Orin Nano 8GB board, where NVIDIA's Cosmos 3 Edge model handles vision while Bolt manages speech replies and tool calls. On benchmark testing conducted by NVIDIA in its own NeMo-Gym harness, Bolt scored 84.8 on IHEval against 67.0 for Gemma 4 E4B and 60.6 for Qwen3.5-4B, and NVIDIA reported roughly 40 percent higher average scores than its own Nemotron 3 Nano 4B. The model does not win every category: Qwen3.5-4B leads on Tau2-Bench (75.1 versus 55.1) and RULER 64K (95.1 versus 87.0), per MarkTechPost's rundown of the vendor-reported figures.

On deployment speed, NVIDIA quantized the model to Q4_K_M with a DFlash speculative-decoding drafter, which reached 272.6 tokens per second on an RTX 5070, about 1.21 times faster than Gemma 4 E4B's own drafter configuration. The model is available now on Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab under the OpenMDW-1.1 license, alongside a free API on build.nvidia.com and packaging as an NVIDIA NIM and through the In-Game Inferencing SDK, MarkTechPost reported.

Why It Matters

Bolt's release lands alongside two other recent robot product stories that illustrate how uneven commercial deployment remains. Wired reported on Mori, a $649 to $849 robot lamp from startup OKA that uses local language models for limited, voice-free interaction, while The Verge detailed how Eiliko maker Energize Lab is hiring an outside security firm after its AI toy exposed customer audio and photos. Cheap, locally-run models like Bolt lower the cost of shipping voice and tool-use capability into consumer and industrial robots without constant cloud dependence, but the Eiliko case shows that shrinking the compute footprint does not resolve data-handling practices around what these devices record. Investors watching the sector for the equity market implications of robotics deployment will want to separate genuine technical capability gains from products still working out basic trust and functionality questions before assuming near-term revenue follows.

What to Watch

NVIDIA's own benchmark claims for Bolt have not been independently verified outside its NeMo-Gym harness, and the gaps on Tau2-Bench and RULER 64K against Qwen3.5-4B suggest tradeoffs rather than a clean win. Watch whether robot and device makers beyond the Reachy Mini demo adopt Bolt in shipping products, and whether NVIDIA's open licensing under OpenMDW-1.1 draws developers away from Apache-licensed alternatives like Gemma 4 E4B and Qwen3.5-4B. The Eiliko and Mori stories also raise the question of whether companies building on models like Bolt will pair local inference with comparable attention to data security and clear feature scope before wider rollout.