Ollama now runs Liquid AI's LFM3 on Linux and Windows, speech included

Ollama now runs Liquid AI's LFM3, small models built for local agents and tool calls, on Apple silicon, Linux and Windows. The text model has 1.7 billion parameters and a 256,000-token context.

Slide text

Ollama

OCT 24–28, 2026

LFM3 runs on Apple silicon, Linux and Windows

1.7B text model

256K token context

Tools tool calling, structured output

$ ollama run lfm3

Ollama release notes, Oct 27 and Oct 28 · PRs #19333, #19344 · ollama.com/library/lfm3

LFM3 VL sees images. LFM3 Audio does speech recognition, translation and text-to-speech in 18 languages, now on Linux and Windows too. LFM3 Omni takes text, images and audio together.

Slide text

Ollama

OCT 24–28, 2026

lfm3-vl image understanding

lfm3-audio speech recognition, translation and text-to-speech in 18 languages

lfm3-omni text, images and audio in one model

Ollama release notes, Oct 27 and Oct 28 · PR #19344 · ollama.com/library/lfm3-audio

Liquid AI's d1 decision models now come in a 2 billion parameter size. Ask one a yes or no question about text or an image. d1 Omni adds audio, and the 2 billion size arrived on Apple hardware first.

Slide text

Ollama

OCT 24–28, 2026

Liquid AI’s d1 decision models now come in a 2B size

$ ollama run d1:2b "I was charged twice for
  order 4417." --ask "Is this a billing issue?"

d1-omni adds audio. The 2B size arrived on Apple hardware first.

Ollama release notes, Oct 27 · PR #19333 · ollama.com/library/d1

You no longer have to write numbers out as words for Phonon. Prices, ranges, units, keyboard shortcuts and web links come out the way a person says them, and markdown symbols are dropped instead of spoken.

Slide text

Ollama

OCT 24–28, 2026

• $4.59 → “4 dollars and 59 cents”
• 5-10 minutes → “5 to 10 minutes”
• Ctrl+C, emails and URLs are said aloud

Ollama release, Oct 24 · PR #19234 · docs.ollama.com/capabilities/speech

On Apple silicon, long prompts get through faster, so replies start sooner. Laguna processes prompts about a quarter faster, Flint 14 percent faster, and Gemma 4 gains 9 percent. Generation speed stays the same.

Slide text

Ollama

OCT 24–28, 2026

+27% Laguna

+14% Flint

+9% Gemma 4

Prompt processing, Apple M5 Max. Generation speed unchanged.

Ollama release notes, Oct 27 · PR #19291

Ollama launch now starts OpenCode version 2 on the model you picked. Kimi CLI tool calls stop failing on turns with no text. And older versions of macOS stop failing with a Metal library error.

Slide text

Ollama

OCT 24–28, 2026

• ollama launch starts OpenCode v2 on the model you picked
• Kimi CLI tool-only turns stop failing
• Older macOS stops failing to build Ollama’s Metal library

Ollama releases, Oct 27 and Oct 28 · PRs #19310, #19312, #19372

For a local agent, I would pull LFM3 first: it is small, it holds a long context, and it runs on Mac, Linux and Windows. Ollama's release notes cover the rest. Which LFM3 model would you pull first?

Slide text

Ollama

OCT 24–28, 2026

LFM3 in the Ollama libraryollama.com/library/lfm3

Ollama release notesgithub.com/ollama/ollama/releases

ollama.com/library/lfm3 · github.com/ollama/ollama/releases