Liquid d1 — small multimodal decision models for the edge
LiquidAI's d1-3B and d1-omni-600M are compact decision models fine-tuned from LFM2.5 bases — images in both builds, audio in the omni one — aimed at edge and on-device deployment with GGUF quantizations. What they are, what they run on, and how they differ from hosted Jev.
What Liquid d1 is
LiquidAI shipped d1-3B and d1-omni-600M in early October 2026, fine-tuned from its LFM2.5-VL-3B Vision-Language and LFM2.5-Encoder-350M bases. The Hugging Face tags read decision, calibration, system-one and edge; the pipeline category is image-text-to-text, meaning the models ingest images as well as text. The d1-omni build extends inputs to audio. The multilingual coverage claimed on the cards is wide: English, German, Spanish, French, Italian, Dutch, Polish, Portuguese, Arabic, Hindi, Japanese, Russian, Turkish, Vietnamese and Chinese.
Small enough to carry
These are edge-sized models: 600M parameters for the omni build, 3B for the larger one. Community GGUF quantizations run through llama.cpp, which makes the omni build a realistic fit for laptops and single-board machines. If you have been looking for a decision model that runs where a GPU does not, this release is squarely aimed at that niche.
What to check before adopting
Three things. The license on the Hugging Face cards is listed as "other", not a standard open license — read its actual terms before shipping a product on it. Calibration is claimed in the tags but the proof is your workload: run your own ten-case validation. And self-hosting means you own the serving, updates and monitoring — small models still need an owner.
How it sits against hosted Jev
Jev is a hosted text decision model: probabilities over your declared options, called with a Bearer key from gateways like Optio. Liquid d1 covers inputs Jev does not — images, audio — and runs where you host it. If the decision reads a picture or an audio clip, d1's niche is real; if it reads text and needs to be thresholded in a hosted pipeline, Jev is the shorter path.
FAQ
- Is Liquid d1 a competitor to Jev?
- Partly — both speak the System One decision vocabulary (the tags on d1's cards say so), but they overlap only on text. d1 distinguishes itself with multimodal inputs (vision, and audio in the omni build) and edge-sized weights; Jev is hosted, text-only and served with keys.
- Can I run d1 on my laptop?
- The d1-omni-600M via llama.cpp GGUF is realistic for Apple Silicon and similar hardware; the 3B build needs more headroom. Check the community quantization threads for tested configurations.
- Is Liquid d1 free?
- The weights download free, but their Hugging Face license field is "other" — read the actual terms before commercial use, and remember your compute and operating effort are the real cost.
- Is this the official LiquidAI site?
- No. Optio is an independent gateway for calling Jev and is not affiliated with TypeSafe, LiquidAI or any model covered here; this page describes d1 in its own words where the facts come from the Hugging Face cards.