NVIDIA has released Nemotron 3.5 Bolt E4B 4GB, an open small language model built for fast conversational agents. Released on October 23, 2026, the model targets voice assistants, game NPCs, and robot interfaces running locally on RTX, Jetson, and Spark hardware — packing agent behavior into about 4GB of VRAM.
The model is post-trained from Google’s Gemma 4 E4B using reinforcement learning only. Notably, NVIDIA skipped supervised fine-tuning entirely: according to the research report, SFT improved tool calling but hurt the model’s conversational quality, so the team used reinforcement learning from verifiable rewards (RLVR) and RLHF through NeMo-RL and NeMo-Gym instead. The datasets and full training recipe are public.
Nemotron 3.5 Bolt E4B 4GB Specs and Performance
The headline numbers: 8B total parameters, 4B effective, with 64K context for conversation history, tool definitions, and application state. It runs on GeForce RTX, Jetson, RTX Spark, and DGX Spark — NVIDIA’s Q4_K_M quantized build used just 3.1 to 4.3 GB of VRAM on RTX in the company’s tests, leaving room for speech and vision models alongside it.
On benchmarks, NVIDIA reports Bolt scored top marks on 15 of 21 benchmarks against Gemma 4 E4B, Qwen3.5-4B, and Nemotron 3 Nano 4B. Its best showing was IHEval at 84.8, far ahead of Gemma 4 E4B’s 67.0 and Qwen3.5-4B’s 60.6. Its weakest was Tau2-Bench at 55.1 — about 20 points behind Qwen3.5-4B’s 75.1.
Built for Voice Agents, Game Characters, and Robots
NVIDIA lists three target workloads: device and voice assistants, interactive characters, and embedded or industrial agents. Bolt combines natural dialogue with multi-turn tool calling, including guidance for asynchronous tool calls. A conversational_mode flag produces short, speech-ready replies, while configurable reasoning and a runtime thinking budget let developers cap latency.
The model also anchors real-world demos: a research report shows a Reachy Mini robot running on a Jetson Orin Nano 8GB, where Cosmos 3 Edge handles vision while Bolt handles speech replies and tool calls. NVIDIA additionally calls Bolt its third-generation small model for NVIDIA ACE game characters.
For deployment, NVIDIA quantized the BF16 checkpoint to Q4_K_M with importance-matrix calibration and shipped a DFlash drafter for speculative decoding. On an RTX 5070, Bolt with DFlash hit 272.6 tokens/s — 1.21x faster than Gemma 4 E4B with its MTP drafter at 225.9 tokens/s.
Also Read: Nvidia Just Made DLSS 5 Twice as Fast With 24% Fewer Parameters and 20% Lower VRAM Usage
Also Read: Microduck Robots Go Viral After Jensen Huang Demo: 20,000 Preorders and Counting for Christmas
Source: MarkTechPost — read the full story there.
