Harvey Tenet shows that a post‑trained legal model can dramatically boost performance on contract drafting, due‑diligence review and precedent search while keeping inference costs stable. The main pain point for law firms and in‑house teams is that Tenet is currently only a research preview—no weights, API or model card are publicly released—so they cannot plug it into their own workflows today.
A practical path forward is to treat Harvey’s published recipe as a blueprint. First, obtain the open‑weight Kimi K3 base (available under permissive licensing). Second, set up a sandboxed reinforcement‑learning environment that mirrors the Legal Agent Benchmark tasks: short partner‑style instructions, a mix of core and peripheral documents, and an expert rubric of atomic pass/fail criteria. Third, run asynchronous rollouts graded by an LLM‑as‑a‑judge, using the reward combination Harvey described (criteria satisfied, legal‑issues solved, all‑pass bonus). Fourth, apply GSPO optimization with a rank‑64 LoRA overlay; the training reportedly used ~150 B300 GPUs for two months, which can be replicated on a comparable cloud GPU cluster or a dedicated on‑premise node.
Because the method rewards shorter trajectories, token consumption drops, translating to lower per‑query cost even after the post‑training overhead. Firms can start with a subset of tasks—such as contract redlining—to validate gains before scaling to full‑scale due‑diligence pipelines. By following the recipe, organizations retain full ownership of the resulting checkpoint, avoid vendor lock‑in, and can continuously adapt the model to new practice areas or jurisdictional nuances.
In short, the obstacle is not capability but accessibility; reproducing Harvey’s Tenet recipe gives law firms a deployable, cost‑effective, and customizable legal intelligence tool they control outright.
#AI #Product #LegalTech #LLM #ReinforcementLearning #Automation