Defense Lawyer Runs Qwen3 Locally for Classified Files, Faces Hallucinations
Logical-Charge8760 · reddit · 2026-07-06
A Brazilian criminal defense lawyer, unable to use cloud APIs due to case confidentiality, deployed Qwen3 35B-A3B (MoE) via llama.cpp on an RTX 4060 (8GB VRAM) + 32GB RAM laptop for legal document adaptation. However, the model frequently hallucinated legal citations with high confidence.
The workflow involved selecting the closest template from a historical document archive, prompting the model to replace factual details, remove inapplicable sections, and preserve legal reasoning and citations. The lawyer is seeking advice on grounding, model selection, and RAG pipeline improvements,while noting that Brazilian Portuguese legal texts pose an additional challenge for small models.
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27