Research note | PLI Labs | June 2026

The method

REAP-style expert pruning observes which routed experts activate under a calibration workload, ranks low-saliency experts, removes the stable droppable core, and renormalizes the router over surviving experts.

For this run, the calibration direction was legal work: long-form synthesis, source-grounded review, legal workflow prompts, structured outputs, refusal behavior, and authored legal reasoning. PLI Labs does not publish the private calibration items or identify the underlying legal source materials.

The main finding

The most important result is not just that a checkpoint was produced. The important result is that Qwen3.5-122B-A10B appears prune-resistant.

The saliency profile did not support a confident aggressive expert cut. PLI Labs therefore published a conservative 0.16 release: each MoE layer goes from 256 routed experts to 216, while retaining the original 48-layer depth and top-8 routing pattern.

This matters because legal AI compression should not be driven by the desire for a clean marketing number. If the model resists pruning, the correct research answer is to say so and preserve the legal behavior that appears most important.

The companion Qwen3.5-91b-a10b-LegalReap-Layerdrop6 model

The Layerdrop6 checkpoint asks a different question. After the conservative 104B expert prune, can depth be reduced in a controlled way?

Removed layers: 8, 9, 12, 13, 16, 17

The companion model reduces the checkpoint from 48 layers to 42 while retaining 216 routed experts per remaining MoE layer, producing an approximately 91B-parameter bf16 checkpoint. That does not prove the smaller model is better. It gives legal evaluators a clean comparison point: base model versus expert-reduced model versus expert-reduced plus depth-reduced model.

What to evaluate next

  • source-grounded summarization with answerable and unanswerable prompts;
  • citation discipline, where every material claim must be supported by a source;
  • refusal behavior when the record is incomplete;
  • drafting and revision quality under lawyer supervision;
  • long-context degradation;
  • hallucinated-authority resistance;
  • quantized serving behavior on the target hardware.

Why publish it

PLI Labs is publishing these models and notes because legal AI should not depend only on closed systems or generic model cards. Lawyers and legal technologists need inspectable artifacts, clear limitations, and model releases that explain what legal behavior the work tried to preserve.

This release is optimistic about the direction and conservative about the claim: Qwen3.5 legal REAP is a serious research candidate for legal-work-oriented model compression, not an autonomous legal advice product.

Release links