Harvey Trains Open Source Models To Encode Law Firm Workflows

2026-06-18 · Source: Artificial Lawyer · Field: Legal & Regulatory — Legal Technology (LegalTech), Corporate Law & Business Legal Services · Depth: Advanced, medium

Summary

Harvey CEO Winston Weinberg confirmed Proof of Concept studies with law firms to train open-source Large Language Models (LLMs) to "encode" their specific workflows for complex matters, including client-specific processes. This initiative aims to apply automation more precisely, mirroring efforts by Kirkland & Ellis, which announced a \$500m AI investment and is reportedly hiring AI infrastructure experts for its own open-source training strategy alongside Palantir. Thomson Reuters is also training open-source LLMs on its vast legal data. This trend marks a return to post-training open-source models, driven by data security, improved performance from specific training, and the integration of agentic flows. Harvey co-founder Gabe Pereyra outlined goals: serving frontier intelligence affordably and securely, and enabling law firms to build and "own their own intelligence" for specialized models. The models will focus on complex client matters spanning months, using agentic systems to control legal tech tools and sub-agents. Harvey has open-sourced benchmarks, showing promising results approaching frontier performance.

Key takeaway

For AI Engineers and Legal Professionals evaluating custom AI solutions, Harvey's approach to training open-source LLMs on firm-specific workflows signals a critical shift. You should explore post-training open-source models to "own your own intelligence" and enhance data security, rather than relying solely on general models. This strategy allows you to encode unique methodologies and client relationships, achieving higher quality automation for complex legal work streams and potentially differentiating your firm's service offerings.

Key insights

Post-training open-source LLMs on proprietary workflows and client data offers superior, secure, and customized legal AI automation.

Principles

• Customization improves general model performance.
• Data security drives on-premise model training.
• Agentic flows enhance specific workflow automation.

Method

Train open-source LLMs on law firm-specific complex client matters, integrating agentic systems to control legal tech tools, sub-agents, and reference data.

In practice

• Encode firm-specific complex work streams.
• Integrate client-specific playbooks.
• Build specialized legal foundation models.

Topics

• Legal AI
• Open-Source LLMs
• Workflow Automation
• Custom Model Training
• Agentic Systems
• Data Security

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Legal Professional, AI Scientist

Related on AIssential

• You NEED to try these open-source AI projects RIGHT NOW — Open-source AI projects offer diverse solutions for enhanced search, local document processing, agentic engineering, and LLM cost reduction.
• Lawyer Uses Claude Skills, Legal World Loses It… — Customizable LLM workflows can encode individual expertise to automate complex professional tasks.
• Custom model training. Domain-specific language models. | Mistral - Mistral AI — Mistral AI enables enterprises to build fully-owned, domain-specialized LLMs through collaborative customization and flexible deployment.
• Harvey Drives Legal Agent Learning Via ‘Harness Engineering’ — Harness engineering and autoresearch significantly boost legal agent performance through iterative self-improvement and environmental feedback.
• The Architecture Behind Open-Source LLMs — MoE architectures and diverse training strategies drive rapid progress in open-weight LLMs, optimizing cost and performance.
• Blueprint First, Model Second: A Framework for Deterministic LLM Workflow — Decoupling LLM decision-making from workflow execution via code blueprints ensures deterministic, verifiable agent behavior.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Lawyer.