Liquid AI's LFM3 models are about 2B parameters, fine-tune on one GPU, and its d1 model answers a list of ticket questions in one pass

Not every step in a workflow needs a frontier model. Liquid AI just released a family built for the small, repetitive ones.

What happened

Liquid AI released LFM3, a set of models in the 2B parameter range that share one base: text, vision, audio, omni and decision models. According to its Hugging Face post:

• The LFM3-1.7B text model fits in about 1 GB at 4-bit precision and has a 256K token context window. Liquid says it scores 38.42 on WebArena-Verified against 3.02 for its predecessor, and is post-trained for agent harnesses like Hermes Agent and OpenClaw.
• LFM3-VL-2B takes images and video. Liquid reports it is on par with Gemma 4 12B across 62 vision benchmarks.
• LFM3-Audio-2B handles speech recognition, speech output and translation across 18 languages, and can turn spoken requests into tool calls.
• The d1 models answer a list of questions about one input, such as a support ticket, invoice or voice note, in a single forward pass, each with a probability. Answers can be yes or no, choices, scores, multi select or values read from documents.

Every model can be fine-tuned on a single GPU, and Liquid publishes notebooks for it. The benchmark numbers are Liquid's own.

My take

The d1 example in the post is a support ticket: is this a refund request, which team, which topics, how urgent. That is the exact step in most of the CRM and helpdesk automations I build, and today it usually goes to a large model with a prompt and a JSON parser.

A model that returns all four answers at once, with a probability for each, is a cleaner fit. The probability is the useful part. You can auto route anything above a confidence line and send the rest to a human review queue, instead of trusting every answer equally.

I would not swap production steps over on benchmark claims alone. Take 200 real tickets a client has already routed by hand, run them through d1, and compare. If it matches on the clear cases and its low confidence scores line up with the messy ones, you have a cheaper, private triage step that can run on the client's own hardware.