Liquid AI d1 (local)

d1-3B and d1-omni-600M are open-weight decision models. They are not chat models: given a state (text, JSON, images, or speech for omni) and named typed questions, they return calibrated probabilities in one forward pass with zero output tokens. OpenTryOn exposes them through the decide service with LiquidD1Adapter (Path B, extra="local").

Good fits: garment attribute classification from photos (neckline, sleeve, pattern), QC gating of generated try-on images, moderation, routing.

	d1-3b	d1-omni-600m
Repo	LiquidAI/d1-3B	LiquidAI/d1-omni-600M
Size	3.12B (bfloat16)	587M (float16)
Input	text / JSON + images	text / JSON + images or one 16 kHz mono speech clip (≤ 30 s)
Context	32,768	16,384
Transformers	>=5.14	>=5.15
Decision Index 0.2.1	48.57	15.95
License	LFM Open License v1.0 (lfm1.0) — review for commercial use	same

Images and audio cannot be combined in one call. The omni card recommends float16 on GPU and advises against bfloat16.

Hosted, text-only twin: Typesafe Jev (--model jev).

Requirements

pip install opentryon[local]
pip install "transformers>=5.15" torchvision pillow soundfile

The shared opentryon[local] pin is transformers==4.42.4, so use a separate environment for d1 (same situation as Limite and Ternary Bonsai). The adapter raises a clear error if the installed transformers is too old.

Optional override: export LIQUID_D1_MODEL_ID=/path/to/snapshot.

CLI

Question schema (noul / choice / score) is documented on the Jev page.

# Image classification
opentryon decide --model d1-3b --questions neckline.json --image dress.jpg

# Text/JSON state
opentryon decide --model d1-3b --questions triage.json --state "Customer says the zip broke after one wear."

# Speech (omni only)
opentryon decide --model d1-omni-600m --questions intent.json --audio request.wav

Flags: --questions/-q (required), --state/-s, --image/-i (one or more), --audio (omni), --variant, --model-id, --device.

Python

from tryon.models.liquid_d1 import LiquidD1Adapter

adapter = LiquidD1Adapter(variant="d1-3B")
out = adapter.decide(
    questions={"neckline": {"type": "choice", "instructions": "Neckline style?",
                            "criteria": {"v-neck": "V-shaped opening", "crew": "round, high"}}},
    image="dress.jpg",
)
print(out["answers"]["neckline"])  # {'choice': ..., 'confidence': ..., 'probabilities': {...}}

Notes

• MCP tools: decide_d1_3b, decide_d1_omni_600m.
• Audio must already be 16 kHz mono; the adapter errors (rather than resampling) otherwise.
• No GPU was available when this was integrated, so weight loading was not exercised end to end — same caveat as the other local adapters.