Model links: Qwen3.5-104b-a10b-LegalReap and Qwen3.5-91b-a10b-LegalReap-Layerdrop6.
What was released
The first artifact, Qwen3.5-104b-a10b-LegalReap, is a conservative expert-reduced checkpoint derived from Qwen3.5-122B-A10B. It reduces each MoE layer from 256 routed experts to 216 while preserving 48 layers, the tokenizer, hidden size, and top-8 routing pattern. The resulting bf16 checkpoint is about 104B parameters.
The second artifact, Qwen3.5-91b-a10b-LegalReap-Layerdrop6, is a smaller companion model. It starts from the 104B REAP checkpoint and removes six decoder layers: 8, 9, 12, 13, 16, and 17. That gives evaluators a second compression axis to test: expert reduction plus depth reduction. The resulting bf16 checkpoint is about 91B parameters.
Why a legal REAP
Most open model compression work is judged against generic or coding-heavy benchmarks. Legal work stresses different behavior. A useful legal assistant has to preserve factual care, source fidelity, drafting judgment, refusal discipline, and workflow awareness.
PLI Labs built this release around that premise. The goal is not to create an autonomous lawyer. The goal is to produce research artifacts for lawyer-supervised evaluation of a frontier-scale open MoE model compressed around legal work.
Why the cut is conservative
The Qwen3.5 run showed a flatter expert-saliency profile than expected. In plain terms: this model resisted obvious aggressive pruning. Instead of forcing a larger compression claim, PLI Labs published the stable 0.16 expert cut and a separate Layerdrop6 derivative.
That makes the release more useful. Legal AI does not need inflated compression headlines. It needs inspectable models, clear caveats, and evaluation paths that legal teams can reproduce.
Public documentation
- Qwen3.5 legal REAP research packet
- Stack and settings summary
- Qwen3.5-104b-a10b-LegalReap model card
- Qwen3.5-91b-a10b-LegalReap-Layerdrop6 model card
- Research notes on the Qwen3.5 legal REAP run
Intended use
These models are research candidates for lawyer-supervised evaluation. They are appropriate for controlled testing of legal drafting, revision, summarization, workflow behavior, and source-grounded analysis. They are not unsupervised legal advice systems, and they are not a substitute for attorney review.
PLI Labs is optimistic about this direction because legal AI should become more private, more inspectable, and more aligned with the actual work lawyers do. The release is conservative because legal reliability has to be earned with evidence.