- Nvidia released Nemotron 3.5 Bolt E4B 4GB, an open small language model post-trained from Google’s Gemma 4 E4B for local voice assistants, game characters, and robots.
- The model targets about 4GB of VRAM and runs on RTX, Jetson, RTX Spark, and DGX Spark hardware.
- Nvidia used reinforcement learning (GRPO) to improve multi-turn tool use, instruction following, and input grounding over the base Gemma model.
- It scored best on 15 of 21 benchmarks against Gemma 4 E4B, Qwen3.5-4B, and Nemotron 3 Nano 4B, including an IFEval score of 84.8 versus Gemma’s 67.0.
What Happened
Nvidia released Nemotron 3.5 Bolt E4B 4GB, an open small language model built by post-training Google’s Gemma 4 E4B for local voice assistants, interactive game characters, and robotics interfaces. MarkTechPost covered the release on Oct. 23, 2026.
Why It Matters
The release continues Nvidia’s push to run capable AI models entirely on local hardware rather than the cloud, complementing its broader infrastructure push, including Nvidia’s DSX POD effort to cut AI data center setup time to 14 days. A model this small running fully on-device matters for latency-sensitive use cases like voice assistants and robots that can’t depend on a stable network connection.
Technical Details
Nvidia post-trained Nemotron-3.5-Bolt-E4B from Gemma-4-E4B-it using multi-environment reinforcement learning with Group Relative Policy Optimization across instruction-following, multi-step tool use, multi-turn conversation, and structured-output tasks. The model targets roughly 4GB of VRAM and runs on Nvidia’s RTX, Jetson, RTX Spark, and DGX Spark hardware. It scored best on 15 of 21 benchmarks tested against Gemma 4 E4B, Qwen3.5-4B, and Nemotron 3 Nano 4B, including an IFEval score of 84.8 against 67.0 for Gemma 4 E4B and 60.6 for Qwen3.5-4B.
Who’s Affected
Developers building voice assistants, game NPCs, or robotics and IoT interfaces get a small open-weight model tuned specifically for multi-turn tool use on consumer and edge hardware. It gives Nvidia’s hardware ecosystem a bundled, optimized model rather than relying solely on third-party fine-tunes.
What’s Next
Nvidia has not announced a roadmap beyond this release, but the post-training approach — taking Google’s base model and tuning it with Nvidia-specific reinforcement learning — suggests further Bolt-series variants could follow as Gemma itself is updated.
