Ollama extends local LFM3 speech inference to Windows and Linux

Ollama extended local LFM3 speech inference to Windows and Linux.

**Dek:** Version 0.40.13 broadens platform support for Liquid AI’s LFM3 family, including audio and omni variants. The release makes it possible to test the same local speech workflow beyond Apple silicon.

Ollama’s October 28 release of version 0.40.13 adds Linux and Windows support for Liquid AI’s LFM3 family. The release notes explicitly include speech generation through the LFM3-Audio and LFM3-Omni models. The previous Ollama release introduced the family in its catalogue, including separate text, vision, audio and omni variants. The platform expansion is the new development. Liquid AI’s model design and Ollama’s ability to run it on a given operating system are separate layers. [Ollama v0.40.13](https://github.com/ollama/ollama/releases/tag/v0.40.13) · [Ollama v0.40.12](https://github.com/ollama/ollama/releases/tag/v0.40.12)

What the audio model does

The Ollama catalogue describes LFM3-Audio-2B as accepting text or audio and producing text, speech or both. It lists multilingual transcription, translation and synthesis across 18 languages and describes speech output at 24 kHz. That gives application developers a model to try for a voice response, a local transcription task or a multilingual assistant. Each task has a different success measure. A transcript can be useful even when synthesized speech would sound unnatural, and understandable output is only one part of a responsive conversation. [Ollama model page](https://ollama.com/library/lfm3-audio)

A product that previously kept its local speech tests on Apple silicon can now add Windows and Linux machines to its evaluation. The tests should record the operating system, accelerator and model variant, along with startup memory and how quickly usable audio arrives. Running a model successfully on one platform establishes compatibility with that setup. It says little about latency on a different driver, GPU or CPU fallback.

Where an offline workflow begins and ends

Running speech inference on the device can keep that stage of a workflow local. A voice product may still use a network service for retrieval, monitoring, tool calls or another model. A deployment that needs to keep recordings private should trace the microphone data and derived text through every stage, including logs and temporary files. The claim worth testing is where the application's data goes, rather than assuming that a local model determines the behavior of the entire stack.

Start with a known recording, a short text prompt and a multi-turn voice interaction on each supported target. Measure recognition errors and audible output separately from playback and interruption handling. The October update adds platforms on which those questions can now be tested with the same model family. [Ollama v0.40.13](https://github.com/ollama/ollama/releases/tag/v0.40.13)