lumen-merged

Lumen estimated ~200× lower inference cost with frontier competitive saferepair

Lumen is a 4b model for automatically verifiable accessibility attribute repair in react and static html. on a matched 43-task development benchmark, Lumen completed 42 safe repairs, compared with 40 for gpt-5.6 sol and 43 for claude opus 4.8.

this is the standalone bf16 edition of Lumen. it includes the qwen3.5 4b base weights with the Lumen lora adapter merged in, so most users need to download only this repository.

the smaller harrrshall/lumen repository is the 252 mb lora adapter. use that edition when you already store the exact qwen base or want to manage adapters yourself.

quick start

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "harrrshall/lumen-merged"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

no peft adapter attachment and no separate base model download are required. the repository is approximately 9.1 gb before local filesystem overhead.

what Lumen does

Lumen generates small structured source patches for automatically verifiable accessibility defects in react and static html. its intended workflow rebuilds the application and validates the patch with browser, keyboard, semantic, visual, regression, and protected path gates.

frontier comparison

provenance

• base model: Qwen/Qwen3.5-4B
• pinned base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
• post training: qlora supervised fine tuning
• merged checkpoint dtype: bf16
• original adapter: harrrshall/lumen

training curve

limitations

Lumen covers deterministic source repair families supported by its evaluator. it does not establish complete wcag or legal conformance. review every patch before merging, and involve qualified accessibility practitioners and assistive technology users for subjective decisions.