Ollama 0.40.12 adds Liquid AI's LFM3 models and a 2B d1 decision model, on Apple hardware first

Ollama 0.40.12 adds LFM3, Liquid AI's new family of small models built for local agents, tool calling and structured output , and a 2B size of Liquid AI's d1 decision models, built on LFM3. In this release both run only on Apple hardware; support for Linux and Windows is described as coming soon. The same release speeds up prompt processing on Apple silicon for Gemma 4, Flint and Laguna and makesollama launchwork with OpenCode v2. The release page dates it 27 October.

What changed

LFM3, a family of small models.The release notes describe LFM3 as built on Liquid's new tri-hybrid architecture and list its members.lfm3is a 1.7B text model with a 256K-token context window ;lfm3-vladds image understanding ;lfm3-audiodoes speech recognition, translation and text-to-speech in 18 languages ; andlfm3-omnitakes text, images and audio in one model. The command the page gives isollama run lfm3.

The text model's library page adds what you need to judge it. The 256K tokens of context count your input and the model's output together. At 4-bit precision it takes about 1 GB of memory, so it fits on any phone or laptop. It is post-trained for agent harnesses — the programs that let a model call tools in a loop — such as Hermes Agent, Pi and OpenClaw. Arabic is among the languages the page lists.

d1 in a 2B size.Liquid AI's d1 decision models now come in a 2B size, built on LFM3. d1 makes decisions about text and images, and d1-omni adds audio. A decision model does not write an answer: it isn't a chat model and doesn't generate text. You give it a state — text, JSON, images or a mix — and a set of typed questions, and it returns a calibrated probability for every answer in a single pass. We explained how these models work in Ollama when they arrived, inOllama 0.35. The release's example asks a yes-or-no question about one sentence: ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"On the model's library page the new tag isd1:2b: 3.9GB to download, with a 125K context window, taking text and images.

Apple hardware only, for now.The release notes are plain about it: lfm3 and d1 2b are supported on Apple hardware in this release, and support for Linux and Windows is coming soon.

Speed and quality on Apple silicon.Prompt processing — the step where the model reads everything you sent before it writes its first word — is faster on Apple silicon for Gemma 4, Flint and Laguna models. Speculative decoding now supports DFlash2 draft models on MLX, Apple's framework for running models on its own chips. In speculative decoding a small draft model guesses the next few words and the large model checks them all at once, which saves time when the guesses are right. Gemma 4 also keeps more detail in small images.

Coding tools. ollama launchnow works with OpenCode v2 , and tool-only assistant turns from Kimi CLI — replies in which the assistant only calls tools and writes no text — are fixed. Under the hood, MLX and llama.cpp are updated.

The release was made from commit f067033, and its changelog line runs from v0.40.11 to v0.40.12 , so everything above is new sinceOllama 0.40.11.

What it means for you

If you have an Apple silicon Mac, update and try ollama run lfm3. About 1 GB of memory at 4-bit precision means it runs beside your other work even on a Mac with little memory. Use it for what it was trained for: tool calling and structured output inside an agent. If you use Hermes Agent, the page gives the commandollama launch hermes --model lfm3. Do not make it your coding agent: the page itself does not recommend it for agentic coding or complex multi-step reasoning, because of its small size , and for questions that need a lot of knowledge it advises retrieval or search.

If you use decision models, name the tag d1:2bexplicitly, as the release's example does. The page's description of plaind1is still the 3B model built on LFM2.5-VL-3B. The 2B's 125K context window leaves room for a long state, such as a whole support thread, before you ask your question.

If you are on Windows or Linux, lfm3 and d1:2bare not supported yet in this release. The OpenCode v2 and Kimi CLI fixes and the updated llama.cpp are not marked as Apple-only, so update for those, and wait for a later release before pulling the new models.

If you start OpenCode with ollama launchand moved to OpenCode v2, this is the release that makes that pairing work.

How to update. Install 0.40.12 from the release page for your system, then open a terminal and type ollama -v: it prints the version you have installed, which should now be 0.40.12. If it shows an older number, quit Ollama completely, start it again and run it once more.

The source

The release page on GitHub, Ollama v0.40.12; and the library pages for lfm3 and d1.