Ollama v0.40.12 Adds Liquid AI LFM3 Family and d1 2B Decision Models
What changed
Ollama released version 0.40.12, introducing support for Liquid AI’s LFM3 architecture and d1 decision models alongside backend performance optimizations for Apple silicon.
Liquid AI LFM3 Model Family
Liquid AI’s new LFM3 models use a tri-hybrid architecture optimized for local autonomous agents, structured outputs, and tool integration. The release brings four variants to Ollama:
lfm3: A 1.7-billion parameter text-only model featuring a 256,000-token context window (ollama run lfm3).lfm3-vl: Extends the base architecture with vision capabilities for image context.lfm3-audio: Provides speech-to-text recognition, text-to-speech synthesis, and audio translation across 18 languages.lfm3-omni: Unifies text, visual, and audio modalities into a single local model.
Liquid AI d1 Decision Models
Built on top of LFM3, Liquid AI’s d1 line consists of 2-billion parameter models designed specifically to evaluate context and make targeted operational choices:
d1: Evaluates decisions based on text and image inputs.d1-omni: Incorporates audio context into decision tasks.
Developers can query d1 models directly from the command line, for example:
ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"Additional System Updates
- Apple Silicon Acceleration: Increased prompt processing speeds on Apple silicon for Gemma 4, Flint, and Laguna models.
- Speculative Decoding: Added support for DFlash2 draft models when running speculative decoding via MLX.
- Vision Refinement: Gemma 4 now retains higher detail when processing small image inputs.
- Integrations and Fixes: Added
ollama launchsupport for OpenCode v2 and resolved an issue causing tool-only assistant turn errors when integrating with Kimi CLI. - Core Dependencies: Updated both MLX and
llama.cppunderlying libraries.
Why it matters
Running agentic workflows locally requires lightweight models that balance long context management with tool-use precision. The 1.7B parameter lfm3 model delivers a large 256K-token context window in a compact memory footprint, offering local developers an alternative to cloud-hosted APIs for long-context tool invocation. While managing local context efficiently remains critical—especially when compared to how cloud tools handle heavy background routines, such as how Claude Code background agent summaries burn tokens—small models designed specifically for tool calls keep hardware resource requirements minimal.
Similarly, specialized 2B decision models like d1 streamline routing tasks, triage logic, and intent classification at the edge without requiring multi-billion parameter generalist models.
What to do next
Developers on supported hardware can pull the base text model using ollama run lfm3 or evaluate classification tasks with ollama run d1:2b.
Caveats and Hardware Support
In Ollama v0.40.12, support for lfm3 and d1:2b is restricted to Apple hardware running macOS. Linux and Windows compatibility for these Liquid AI models is currently in development and expected in an upcoming release.
Sources
Share Article
If this article helped you, please share it with others!