Research note | PLI Labs

The problem

Most public model-compression work is measured against coding or general chat. That is useful research, but it can preserve the wrong parts of a model for law practice. Legal work stresses long factual synthesis, careful treatment of sources, formal advocacy, and a conservative habit of refusing when the record is not strong enough.

What a REAP is

Many leading open models are mixture-of-experts systems. Only some experts activate for a given workload. A REAP-style prune observes which experts legal workloads use, then removes experts that barely contribute to that legal behavior. The goal is a smaller checkpoint that is easier to serve locally while preserving the legal skills that matter.

Plain English: PLI Labs is not trying to make a model smaller at any cost. The point is to keep the lawyer-facing capability and remove capacity that the legal workload does not use.

The MiniMax-M2.7 campaign

The public MiniMax-M2.7 notes describe two benchmark-backed pruning candidates. The 0.35 lane is the conservative candidate, with 167 routed experts remaining per prunable layer. The 0.45 lane is the more aggressive candidate, with roughly 141 experts remaining per prunable layer. Both were calibrated around legal drafting and harness-style behavior instead of coding benchmarks.

The current public status is intentionally cautious but now release-oriented: PLI Labs is publishing formal model cards, Hugging Face release pages, and GGUF artifacts for the 0.35 and 0.45 legal REAP lanes while continuing to label the work as lawyer-reviewed research rather than autonomous legal advice.

Why lawyers should care

Local legal AI is constrained by hardware, cost, context length, and confidentiality. A properly validated legal prune could let a firm run stronger legal drafting and analysis models on owned hardware. An improperly validated prune could remain fluent while quietly losing the exact judgment lawyers need. That is why the public PLI position is validation first, release second.

Source links