- Home
- PLI Labs
PLI Labs — the open arm
Open research, so you can check our work.
PLI Labs publishes expert-pruned legal model checkpoints, serving recipes for owned hardware, and open-source tools for legal document work. Download them, inspect them, run them on your own machines — no account, no API, no meter.
The method
Prunes, not finetunes.
The strongest open models are mixture-of-experts models: inside one large model live many specialist sub-networks, and only a few activate for any given passage. Run such a model on a steady diet of legal work — pleadings, discovery, financial records, case analysis — and a pattern emerges: some experts do the heavy lifting and many barely fire.
A legal prune measures which experts legal work actually uses and removes the rest. What remains is a smaller, faster checkpoint of the same model that fits more modest hardware. The surviving weights are the original weights; nothing is retrained on new data, which is why a prune is not a finetune — and why no client data is involved.
Every release ships with a model card describing the base model, the pruning recipe, the evaluation, and known limitations, so the results can be reproduced or challenged.
Releases & research
Published in the open.
- Hugging Face · checkpoint
MiniMax-M2.7 legal REAP
Expert-pruned legal checkpoints at the 0.35 and 0.45 pruning lanes, with formal model cards.
- Hugging Face · GGUF
Quantized GGUF for owned hardware
A quantized artifact of the pruned model for local evaluation on accessible GPUs.
- Research note
What the Qwen3.5 run showed
A candid negative result: the model proved prune-resistant, so Labs published a conservative cut rather than a bigger headline.
- Explainer
Legal REAPs, not finetunes
Why Labs prunes open models around legal capability instead of training on client data.
- Owned hardware
Serious legal AI on V100 hardware
Why topology, backend, and cache settings decide whether an older GPU fleet is useful.
- Owned hardware
A 122B-class legal lane on four B70 cards
A reliable four-card serving lane, measured results, and the guardrails around it.
- Recipes
B70 and V100 launch profiles
Measured launch settings and token rates for owned inference lanes.
- Open-source tools
Fail-closed legal AI tooling
Tools that stop, record the reason, and require review when the evidence is weak.
Why publish
Claims you can test yourself.
Lawyers are asked to trust AI systems they cannot inspect. We think the profession is better served by work it can download, examine, and run.
Publishing also keeps us honest. Our claims about what works for legal AI are testable by anyone with a graphics card, and the methods our clients rely on have survived public scrutiny. The two arms feed each other: consulting and a working litigation practice tell Labs what to build, and Labs gives legal organizations a no-cost way to evaluate our work before hiring anyone.
Legal-technology teams and model builders: we collaborate on legal evaluation sets, workflow design, and open legal models. Get in touch.