NVIDIA has launched Nemotron 3.5 Bolt E4B 4GB, an open small language model designed for fast, local conversational agents. Post-trained exclusively via reinforcement learning (RLVR and RLHF) from Google’s Gemma 4 E4B, the model targets low-resource hardware like RTX, Jetson, and Spark. Skipping supervised fine-tuning allowed the team to preserve dialogue quality while improving tool calling.

The model operates within a 4GB VRAM envelope, leaving operational headroom on edge devices for concurrent vision and audio models. In hardware demonstrations, NVIDIA showcased the model driving speech and tool invocation on a Jetson Orin Nano alongside vision models. To achieve real-time performance, the checkpoint was quantized to Q4_K_M using importance-matrix calibration and paired with a spec-decoding drafter, delivering up to 272.6 tokens per second on an RTX 5070.

While the model outperforms predecessors like Nemotron 3 Nano 4B and leads in specific multi-turn benchmarks, competitive open models such as Qwen3.5-4B still maintain advantages in long-context and specific agent benchmarks. Bolt is available immediately across open repositories, Ollama, LM Studio, and as an NVIDIA NIM.

Why it matters

  • Enables local execution of complex conversational and tool-calling agents alongside vision models within consumer-grade 4GB VRAM envelopes.

  • Demonstrates that pure reinforcement learning post-training without SFT can yield superior multi-turn tool calling without degrading conversational fluency.

  • Provides optimized deployment pathways for edge AI, robotics, and game NPCs via llama.cpp, Ollama, and NVIDIA NIM integration.

Source: marktechpost.com