Public legal AI projects from Proprietary Legal Intelligence.
PLI Labs publishes legal-specific model work and legal AI tooling for lawyers who want systems they can inspect, run locally, and evaluate against real legal workflows.
The core idea is simple: most open model compression and pruning work is aimed at coding. Legal drafting, long-form analysis, source-grounded summarization, and judgment-heavy document work activate different capabilities. A legal model release should explain what legal workload shaped it, what was intentionally preserved, what was allowed to degrade, and what evidence supports the result.
Start here:
- wiki/README.md - public PLI Labs wiki and release thesis.
- reaps/README.md - what a legal REAP is and why it differs from coding-oriented REAPs.
- huggingface/README.md - public Hugging Face release documentation and model-repository links.
- research/qwen35-legal-reap/README.md
- Qwen3.5 legal REAP research packet, public settings, and artifact inventory.
- blog/proprietarylegal/README.md - public blog prose for ProprietaryLegal.com.
- finetunes/README.md - how PLI thinks about legal finetunes after pruning.
- tools/README.md - public tooling principles for legal AI pipelines.
- research/v100/README.md - V100 inference research, public-safe benchmark summaries, and testing standards.
- research/b70/README.md - Intel Arc Pro B70 inference research, sanitized stack findings, and legal-context benchmark bands.
- PLI Labs is distinct from private law-practice automation and client work.
- Public-release work should be separable from private client materials, attorney-client communications, and privileged work product.
- Do not add private client files, closed-client materials, credentials, tokens, or machine-local secrets to this repo.
The MiniMax-M2.7 0.35 and 0.45 legal REAP materials are published as benchmark-backed candidate notes with Hugging Face release documentation.
The Qwen3.5-104b-a10b-LegalReap and Qwen3.5-91b-a10b-LegalReap-Layerdrop6 artifacts are public research candidates derived from Qwen/Qwen3.5-122B-A10B. The expert-reduced checkpoint is about 104B parameters; the Layerdrop6 companion is about 91B parameters. Public aggregate legal benchmark tables should be added before broader validated-release claims.
- Website: https://proprietarylegal.com
- Research site: https://proprietarylegal.ai