Skip to main content
← Back to Decision Models
This model is deprecated. Use d1-2B for improved accuracy with a third fewer parameters, plus typed answers.
d1-3B is a 3.1B parameter decision model built on LFM2.5-VL-3B. It evaluates text, JSON, images, or a mix and returns calibrated Noul, Choice, and Score answers in one forward pass with zero output tokens.

Specifications

Multimodal

Text and images in the same state

Zero Output Tokens

Answers read from the distribution, no generation

Fast

Single forward pass, no autoregressive generation

Quick Start

Install:
Text input:
Image input:

Other Formats

  • MLX: d1-3B-MLX-8bit, d1-3B-MLX-4bit, and d1-3B-MLX-fp16 run on Apple Silicon with mlx-vlm. They require mlx-vlm 0.7.8 or later, installed with pip install -U "mlx-vlm>=0.7.8". See the model cards for Python, CLI, and server usage.
  • W8A8: d1-3B-w8a8 is an INT8 (weights and activations) version for NVIDIA GPUs with INT8 tensor cores, from Ampere on, including Jetson devices. It needs transformers>=5.19 and torchao>=0.18, and is 3.8 GB instead of 6.2 GB for the bf16 checkpoint. On other hardware, use d1-3B. See the model card for setup.
Last modified on November 17, 2026