Ollama 0.40.10 adds Nemotron Bolt E4B support and DFlash speculative decoding

Ollama released version 0.40.10 on October 23. Its main update is support for NVIDIA Nemotron 3.5 Bolt E4B, released the same day, along with a DFlash draft model for speculative decoding. Developers can launch it with ollama run nemotron-3.5-bolt:e4b. Following the previous version’s voice features, this is another update focused on running models locally. Release announcement

Bolt E4B is post-trained from Gemma 4 E4B. It has about 8 billion total parameters and about 4 billion effective parameters, and is designed for use cases including multi-turn tool use, instruction following, and device assistants. NVIDIA’s model card says training used multi-environment GRPO reinforcement learning across tool use, conversation, and structured output. It will be worth seeing whether the model can maintain coherent agent workflows and output formats on limited hardware. NVIDIA model card

DFlash provides an additional draft model for speculative decoding: the draft model proposes candidate content, which the main model then verifies. Ollama’s model page lists it as a companion for faster generation and says E4B supports Windows, Linux, and Apple Silicon. The release announcement, however, provides no comparable speed figures for these platforms. Measure the actual gains with your own prompt lengths, tool responses, and hardware. Ollama model details

Take care to distinguish the packages when deploying. Ollama’s e4b tag currently lists a size of about 7.2–7.9GB, a 128K context window, and text/image input. NVIDIA’s “4GB” Q4_K_M model card, by contrast, lists text input and output, with up to 64K tokens for each. These specifications are not interchangeable, and the weight size alone does not represent the memory requirements of the full serving setup. Ollama package details, quantized model card

For Chinese-language agent applications, the next step is to test Traditional Chinese tool arguments, instruction retention in long conversations, and changes in latency and memory use with the draft model enabled. Available sources confirm the model’s release and runtime support, but they are not enough to infer its quality or speedup on Chinese workloads.