d1-2B-MLX-4bit
4-bit MLX version of d1-2B for Apple Silicon. The weights take 1.35 GB.
d1-2B is a decision model for text, pictures, and video. It answers named, typed questions over a state with calibrated probabilities in one forward pass, without generating tokens.
The language model uses 4-bit affine quantization with groups of 64. The vision tower, projector and norms stay in FP16. Against the original checkpoint run in float32, it changed 3 of 52 test answers and moved probabilities by up to 0.25. Prefer FP16 or 8-bit where calibration matters.
Find architecture and usage details in the original model card.
Use with MLX-VLM
Requires MLX-VLM 0.7.11 or later:
pip install -U "mlx-vlm>=0.7.11"
d1-2B answers questions without generating text, so chat apps and mlx_vlm.generate cannot run it. Use model.system_one or the decide command below.
from mlx_vlm import load
from mlx_vlm.utils import load_image
model, processor = load("LiquidAI/d1-2B-MLX-4bit")
# Text: several named questions over one state, answered in one pass
ticket = "I was charged twice and my parcel never arrived. Please refund one charge."
out = model.system_one(ticket, {
"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
"team": {"type": "choice", "instructions": "Which team should handle this?",
"criteria": {"billing": "Charges, refunds, invoices",
"shipping": "Delivery and parcels",
"technical": "App or site faults"}},
"topics": {"type": "choice", "multiple": True,
"instructions": "Which topics does the message raise?",
"criteria": {"billing": "Payments and charges",
"delivery": "Shipping and parcels",
"login": "Signing in to the account"}},
"urgency": {"type": "score", "instructions": "How urgent is this ticket?",
"criteria": ["low", "medium", "high"]},
})
a = out["answers"]
print(round(a["refund"]["noul"], 2), a["team"]["choice"], a["topics"]["choices"],
a["urgency"]["legend"][str(int(a["urgency"]["score"]))])
# 1.0 billing ['billing', 'delivery'] medium
print(out["usage"])
# {'input_tokens': 189, 'output_tokens': 0}
# Picture: the photo is the whole state
image = load_image("http://images.cocodataset.org/val2017/000000039769.jpg")
out = model.system_one(None, {
"cats": {"type": "number", "instructions": "How many cats are there?", "range": [0, 10], "step": 1},
"remote": {"type": "box", "instructions": "Where is the remote control on the left?"},
}, media=[image])
a = out["answers"]
print(a["cats"]["number"], {k: round(v, 2) for k, v in a["remote"]["box"].items()})
# 2.0 {'top': 0.16, 'bottom': 0.24, 'left': 0.06, 'right': 0.28}
model.system_one returns {"answers": {name: answer}, "usage": {...}}, and model.system_one_batch takes several requests. The original model card describes every question type and answer field.
From the command line, the answers print as JSON:
python -m mlx_vlm decide --model LiquidAI/d1-2B-MLX-4bit \
--state "I was charged twice this month. Please refund one of the charges." \
--questions '{"refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}}'
More options in the MLX guide.