Liquid AI LFM3: six 2B open-weight models, from text to decisions

Liquid AI says its LFM3 family brings text, vision, audio, speech-and-vision ("omni") and decision models onto one 2-billion-parameter-class backbone, all open-weight on Hugging Face. The company announced it in a post on the Hugging Face blog dated 27 October 2026. The post is vendor-published, and every benchmark below is Liquid AI's own.

What is in the release

• LFM3-1.7B (text): Liquid AI says it outperforms the earlier LFM2.5-2.6B with 35% fewer parameters, fits in about 1 GB at 4-bit precision, and has a 256K-token context window. It is post-trained for agent harnesses such as Hermes Agent, Pi and OpenClaw.
• LFM3-VL-2B (vision, now with video input), LFM3-Audio-2B (speech in 18 languages) and LFM3-Omni-2B (text, images, video and audio in; text and speech out).
• d1-2B and d1-omni-2B: models that answer a list of questions about a ticket, invoice, image or voice note in a single forward pass, each answer with a calibrated probability, rather than generating text.

The numbers, and who ran them

On WebArena-Verified, a browser-task benchmark, Liquid AI reports LFM3-1.7B scores 38.42 against 3.02 for LFM2.5-2.6B, and says it beats Qwen3.5-9B on long browser and business-app tasks. For vision, it reports an average across 62 benchmarks of 71.4, up from 57.6 for its previous 3.1B model, on par with Gemma 4 12B. For the d1 decision model it reports 75.5 on ImajevBench, against 64.0 for the d1-3B it replaces. These are the vendor's measurements; the post does not say any were reproduced independently.

Why a team running AI at work might care

The decision models are the most directly useful idea. Routing a support ticket, flagging a refund request or scoring urgency is a classification job. The post's own example asks one question set of a ticket (refund requested, which team, which topics, urgency) and returns typed answers with probabilities from one pass, which can be thresholded and logged. That is easier to audit than free text.

Liquid AI also lists fine-tuning notebooks that each run on a single GPU, including extracting structured JSON from messy documents and reading receipt photos into JSON. The pitch is that a small model tuned on your data, running on your own hardware, can replace a general model call for narrow tasks.

What the post leaves out

• No licence terms appear in the post. It says the models are open-weight; check each model card on Hugging Face before shipping commercially.
• No cost comparison, latency table on a named server, or customer deployment is given. The one applied example, a robotics stack, reports 96.2% average success on the LIBERO simulation suites, not a production workload.
• The ticket-routing example is a code snippet, not an accuracy result on real support data. Fine-tune results on your own labels are the number that matters, and nobody has published it for you.