NVIDIA releases Bolt E4B weights for local assistants
NVIDIA has opened downloads for Nemotron 3.5 Bolt E4B, a text model post-trained from Gemma 4. It supports tool calls across conversations, with separate controls for thinking and conversational output. Compact weights and a full-precision version are offered in separate repositories. The intended uses include local assistants and interactive characters, with application-specific testing required before integration.
Artificial Intelligence··Midday
Bolt’s downloadable weights are available
NVIDIA made Nemotron 3.5 Bolt E4B, its model for local assistants, downloadable on October 23. The previously announced model now has weights available to developers.[1], [2]
Bolt was post-trained from Google’s Gemma 4 E4B and takes text input and produces text output. Its model card documents no native image or audio input. Intended applications include device and voice assistants, game characters and embedded interfaces. Multi-turn tool calling is designed to support application actions during an ongoing conversation.[1]
Thinking and conversation have separate controls
A chat-template setting switches thinking on or off, while a runtime budget can constrain reasoning. A separate conversational mode controls conversational output. The stated goals also include instruction following and grounding answers in supplied context. Post-training used reinforcement learning across multiple task environments with NVIDIA’s NeMo RL and NeMo Gym tools. The May training-data cutoff is distinct from October’s public availability.[1]
The target and draft models are separate
The compact checkpoint uses Q4_K_M quantization, while full-precision BF16 weights are in a separate repository. A separate DFlash draft model supports speculative decoding, proposing output for the target model to verify. Deployment examples run the target alone or alongside that helper. Commercial use is governed by OpenMDW-1.1. Before application integration, the model card calls for unit- and system-level validation with data specific to the intended use.[1]
Related columns
For more information on this topic, you can read the related columns.
