LFM3-VLD-2B

Developers can use it to generate text from prompts and make decisions about browser tasks, including tasks involving images.

Availability

Model checkpoint available on Hugging Face with Transformers quickstart instructions.

Pricing model

Not established

Deployment

Transformers quickstart; companion GGUF and WebGPU-ready ONNX repositories.

Privacy

Not established

Licence

license:other

• The model combines a 24-layer hybrid decision backbone, a vision tower, a multimodal projector, and embeddings from LiquidAI/d1-2B. ^[1]

  **Decision Backbone**: The complete 24-layer hybrid architecture (Conv + GQA attention + Feed-Forward), vision tower, multimodal projector, and embeddings from `LiquidAI/d1-2B` (trained on browser actions, verification, and decision trajectories).

• The model uses the original language head from LiquidAI/LFM3-VL-2B, untied from token embeddings. ^[1]

  **Restored Language Head**: The original uncorrupted generative language head (`lm_head.weight`) from `LiquidAI/LFM3-VL-2B`, untied from token embeddings (`tie_word_embeddings: false`).

• The publisher reports 91.3% accuracy for the Decision Probe on the d1 Balanced Decision Benchmark. ^[1]

  - **Preserved decision separability**: Hidden states retain deep decision knowledge, reaching **91.27% accuracy** with **0% false completions** on the balanced browser decision benchmark.

• A Transformers quickstart uses AutoModelForImageTextToText and AutoProcessor with trust_remote_code enabled. ^[1]
• The model is published at webbrain-one/lfm3-vld-2b on Hugging Face.