SaulLM 7B

SaulLM-7B (paper: March 2024, arXiv 2403.03883) was the first large language model built explicitly for law. Researchers from Equall.ai, Université Paris-Saclay, Sorbonne, and Lisbon universities continued-pretrained Mistral-7B on 30B+ tokens of legal corpora from the US, UK, Canada, and Europe, then instruction-tuned with legal-specific data — releasing base and instruct variants under the MIT license, plus a LegalBench-Instruct evaluation protocol.

It demonstrated state-of-the-art legal document understanding for its size at release and catalyzed the legal-LLM subfield (followed by SaulLM-54B/141B and many domain successors).

Honest limits: 2024 vintage and 7B scale — modern generalist models often match it on legal tasks; trained overwhelmingly on English-language common-law sources (other jurisdictions/languages underrepresented); hallucination risk is acute in law (fake citations have caused real court sanctions) — never rely on it for filings without verification.