Liquid AI's "judgment model," post-trained on LFM3-Omni-2B, is the d1-2B version with added audio input: given a text or JSON state, images, video frames, audio clips, and several named questions (which can be written or spoken), it returns the answer and probability for each question in a single forward pass without generating text; it is designed for call and voicemail routing, acoustic event recognition, and the same text and image judgments as d1-2B.
Upvotes 81 · Likes 81 · Downloads 1083
Reasons
- Question types cover yes/no judgments, single/multiple choice, 2–10 level scoring, extracting text/numbers/dates from state (with source positions), quantity estimation with p10–p90 intervals, and bounding one or all objects in an image.
- Officially claimed to perform each judgment in 5.8 ms on H100, 3.7 ms on AMD MI325X, 12.4 ms / 8.2 ms with 10 seconds of audio, and 21.1 ms on Apple M5 Pro.
- Official audio tests (effective accuracy, 0 is random): SLURP listening for human voice assistant command intent 81.7, ESC-50 environmental sounds 84.5, 43 language identification 92.5; the d1-2B version received on 29 10 could not process audio and was not included in previous issues.
Usage
- pip install "transformers>=5.19" torch torchvision pillow; AutoModel.from_pretrained("LiquidAI/d1-Omni-2B", trust_remote_code=True), call model.system_one(state, {question_name: {type, instructions, …}}, media=[{"type": "audio", "path": "call.wav"}]); no sampling parameters are adjustable.
- Audio uses 16-bit PCM WAV (automatically resampled to 16 kHz); for spoken questions, replace instructions with an audio clip; for batching, use system_one_batch, first warm up with model.compile() (initial compilation takes about one minute).
- Missing values or boxes return None when found is below 0.5; judgment thresholds should be calibrated with your own business data.
Risks
- It is not a chat model and does not output transcriptions or speech; accuracy drops significantly for long spoken passages (Belebele listening 41.8, written 83.6), especially for non-Latin script languages; speech does not support Hindi, Thai, Vietnamese, or Indonesian.
- When audio is not needed, the official recommendation is to use d1-2B: it is stronger in text and single-box tasks and has better calibration; this model's average public benchmark score is 67.7, lower than d1-2B's 70.2.
- The license is LFM Open License v1.0, not Apache / MIT; trust_remote_code is required; all scores are official self-tests.