NVIDIA releases Nemotron 3.5 Bolt E4B, a 4GB open model for local agents

NVIDIA announces Nemotron 3.5 Bolt E4B 4GB, an open small language model built on Gemma 4 E4B with Nemotron post-training, aimed at conversational agents on local hardware. The report says it offers about 40% higher average quality than NVIDIA's prior model, Nano 4B, on conversational agent benchmarks, and up to about 80% fewer output tokens than peers. It supports multi-turn tool use, a conversational mode, configurable reasoning, a 64k context, and 35 languages. Performance figures are reported across RTX, Jetson, and DGX Spark hardware using Q4_K_M quantization.

Key takeaways

• Bolt E4B reports about 40% higher average benchmark scores than the prior Nano 4B model across the cited conversational and agentic benchmarks.
• The model supports multi-turn tool use, a conversational mode for speech-ready replies, and configurable reasoning with a 64k context window.
• Reported single-session results include about 15.5 tokens/s on Jetson Orin Nano 8GB and 133.4 tokens/s on RTX 5070 at 3K input, 50 output.
• Benchmark results come from NVIDIA's own evaluation setup and may vary with hardware, inference software, and configuration.

Why it matters

Small, local models that handle dialogue and tool calls within a few gigabytes of memory could enable responsive assistants on consumer and embedded devices without cloud latency or privacy tradeoffs.

Read with care

Source claims not independently checked; benchmark comparisons are reported by NVIDIA and some evaluations, including Brevity Bench and When2Call-MT, were noted as to be released later.