REDDIT · REDDIT// 21d agoTUTORIAL

Local Qwen 3.5 agent transcribes 1938 sci-fi PDF

A developer successfully transcribed a poorly-scanned 1938 sci-fi PDF using a local Qwen 3.5 27B agent and Llama.cpp. The model recovered lost spaces and line breaks, producing a readable short story after initial tool-calling hiccups.

// ANALYSIS

Running a 27B model for document transcription is compute-heavy but effective when traditional OCR fails on formatting.

–The user ran Qwen 3.5 27B locally across a 3090 and 3060, achieving ~16 tokens/second at 120k context
–Local agents often struggle with proper tool-calling syntax (e.g., malformed JSON when chunking large outputs)
–Demonstrates how open-weight models can act as intelligent text-restoration tools for archival documents
–Highlights the ongoing friction of local agent frameworks, requiring manual intervention for bash tool usage

// TAGS

qwen-3.5llama-cppagentllminferenceopen-weights

DISCOVERED

21d ago

2026-03-22

PUBLISHED

21d ago

2026-03-22

RELEVANCE

6/ 10

AUTHOR

neph1010