LFM3-VLD-2B

LFM3-VLD-2B (Restored Language Head) is a hybrid multimodal browser decision model built using the Pegasus method (inspired by AveLabs/Qwen3.8-27B-Perplexing-Pegasus).

It combines:

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).
2. 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 Problem Solved

When models like LiquidAI/d1-2B undergo full fine-tuning on decision tasks with tied word embeddings, the language head weights drift away from fluent language generation, causing token repetition and degraded prose in standard causal LM generation. By untying the embeddings and restoring the original LFM3-VL-2B language head onto the decision-trained backbone, LFM3-VL-2B-VLDR achieves:

• Flawless open-ended multimodal chat & text generation (no repetition artifacts).
• Preserved decision separability: Hidden states retain deep decision knowledge, reaching 91.27% accuracy with 0% false completions on the balanced browser decision benchmark.

Benchmark Results (d1 Balanced Decision Benchmark — final_test, 126 Categorical Decisions)

Companion Repositories

• webbrain-one/lfm3-vl-2b-vldr-gguf: GGUF quantizations (BF16, Q8_0, Q4_K_M, mmproj-F16) for llama.cpp (/v1/chat/completions + /v1/systemone).
• webbrain-one/lfm3-vl-2b-vldr-onnx: WebGPU-ready ONNX bundle (decoder_q8f16, decoder_q4f16, vision_encoder_fp16, and decision probe JS runtime).

Quickstart

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "webbrain-one/lfm3-vl-2b-vldr"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "What is the capital of France?"}
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, return_tensors="pt").to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=64)
print(processor.decode(out[0], skip_special_tokens=True))
# The capital of France is Paris.