Ollama 0.40.13 Extends LFM3 Support to Linux and Windows, Including Speech Generation Models

Ollama released version 0.40.13 on October 28, extending support for Liquid AI’s LFM3 series to Linux and Windows, and explicitly covering the text-to-speech features of lfm3-audio and lfm3-omni. This is a deployment update following the models’ release: the execution path previously available on Apple silicon now extends to two additional operating systems. Official release notes

LFM3-Audio-2B accepts text or audio input and can output text, speech, or interleaved text and speech. The model combines a causal audio encoder with the LiquidCodec audio tokenizer. Its specifications list 16 kHz audio input, 24 kHz speech output, incremental prefill, and streaming speech generation. The language list includes Chinese, and the model can also turn spoken requests into tool calls. Audio model details

LFM3-Omni-2B adds vision integration. With approximately 2.17 billion parameters, it can understand text, images, video, and audio, then respond with text or speech. The stated goal is to use a single model for tasks traditionally handled separately by speech recognition, vision-language models, language models, and speech synthesis. Video is sampled at two frames per second, so that sampling limit should be considered when evaluating fast motion or brief events. Omni model details

For engineering teams, this update expands the environments available for testing local multimodal applications. Based on the capabilities described above, teams can evaluate prototypes for voice assistants, audio and video question answering, and device control on Linux workstations or Windows PCs. Whether this reduces the maintenance cost of connecting multiple models still needs to be validated on real tasks; it cannot be inferred from the support announcement alone.

Two limitations should be checked before deployment. The specification tables for both Audio and Omni list a 32K context window, while the labels near the top of the Ollama pages show 125K. These figures conflict, so capacity planning should not rely on the larger number without verification. In addition, Omni explicitly does not support thinking, and Audio is not recommended for complex, multi-step reasoning. The release announcement provides no hardware compatibility matrix or latency figures. Teams should measure memory use, time to first speech segment, and Chinese output quality on their target machines. Audio specifications, Omni specifications