Evergreen Tools
|
← Back to AI Tools

LFM3: Text, Vision, Audio and Decision Models on One 2B Backbone

Liquid AI announced LFM3 on 27 October 2026: text, vision, audio, omni and decision models share one new backbone, the first release sits in the 2B range, and the base plus post-trained weights are open on Hugging Face

Tool Interface

Interactive tool will be available soon

Features

  • ✓ Tri-hybrid layers: of 24 layers, 13 are gated short convolutions, 5 are sliding-window attention with a 256-token window and 6 are global attention, with the convolution kernel widened from 3 to 7
  • ✓ Pre-trained on 22T tokens, with the final 3.5T a curated mid-training stage weighted toward agentic traces and diverse tool calls
  • ✓ Cheaper long context: Liquid says LFM3-1.7B matches LFM2-1.2B on short prompts on edge devices and runs roughly 25% faster at 8K tokens of context
  • ✓ Wide device coverage: Liquid validated 14 device and backend combinations including Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, Arm and x86 cloud CPUs and the Qualcomm Hexagon NPU
  • ✓ A 128K vocabulary that Liquid says tokenizes non-Latin scripts more efficiently

How to Use

  1. Pull weights from the LiquidAI/lfm3 collection on Hugging Face, or try the official Playground and API first
  2. Run inference locally per docs.liquid.ai and measure first-token latency and throughput on your actual target device
  3. Fine-tune on your own narrow task, which is exactly what Liquid pitches this family for
  4. Re-run the official Pipette benchmarks before swapping out a production model

FAQ

When did it launch, and are the weights open?

27 October 2026. Official blog: https://www.liquid.ai/blog/lfm3 . Both base and post-trained models are open-weight and available on Hugging Face, the Liquid Playground and the Liquid AI API.

How big are the models?

Liquid says the first release covers the 2B parameter range, with LFM3-1.7B as the text model, alongside vision, audio, omni and decision models on the same backbone. Per-variant parameter counts and quantisation live on the Hugging Face model cards.

What actually changed versus LFM2?

Two architectural moves: a new tri-hybrid mix of gated short convolutions, sliding-window attention and global attention, and a much larger hardware-in-the-loop search, from one phone plus one laptop for LFM2 to 14 device and backend combinations for LFM3. Liquid claims more capability per parameter and about 25% faster long-context inference.

Can I fine-tune it locally?

That is the pitch: fine-tunable locally, runnable on any device, capable enough to become a reliable expert on one specialised task. Running and fine-tuning steps are in docs.liquid.ai.

What about the licence?

The blog states the weights are open; the binding terms are on the Hugging Face model cards. Read the card, the licence file and your deployment shape (local versus hosted) together before shipping.