Qwen3.5-91b-a10b-LegalReap-Layerdrop6 is the depth-reduced companion to PLI Labs' conservative Qwen3.5 legal REAP checkpoint. It starts from the REAP-0.16 model and removes six low-impact decoder layers selected by a calibration-driven layer-importance pass.

This release explores a second compression axis. REAP reduces routed expert capacity; Layerdrop6 reduces depth. The result keeps 216 routed experts per remaining MoE layer and reduces the network from 48 layers to 42 layers.

This is a research/evaluation release, not an unsupervised legal advice system. Legal use requires attorney supervision and independent source verification.

Release Positioning

The Layerdrop6 model is intentionally presented as a companion artifact, not as a replacement for the 48-layer REAP-0.16 checkpoint. It is for evaluators who want to compare legal behavior across:

the original Qwen3.5-122B-A10B base model;
the conservative expert-reduced REAP-0.16 checkpoint;
a smaller depth-reduced derivative of that checkpoint.

Model Details

Field	Value
Model family	Qwen3.5 MoE
Publisher	PLI Labs / Proprietary Legal Intelligence
Release status	Research candidate
Parent checkpoint	ProprietaryLegal/Qwen3.5-104b-a10b-LegalReap
Original base model	Qwen/Qwen3.5-122B-A10B
Base license	Apache-2.0
Format	Transformers/safetensors checkpoint
Architecture	Qwen3_5MoeForCausalLM
Compression method	REAP expert reduction plus whole-layer drop
Dropped layer indices	8, 9, 12, 13, 16, 17
Hidden size	3072
Layers	42
Parent layers	48
Routed experts	216 per remaining MoE layer
Original routed experts	256 per MoE layer
Experts per token	8 routed experts
Attention heads	32
Key/value heads	2
Head dimension	256
Vocabulary size	248,320
Indexed weight entries	27,855
Indexed tensor bytes	182,327,694,336
Estimated total parameters	about 91B
Contact	nick@proprietarylegal.com

Intended Use

This candidate is intended for:

legal-domain compression research;
side-by-side evaluation against the 48-layer REAP-0.16 model;
lawyer-supervised legal drafting and synthesis experiments;
source-grounded summarization and review experiments;
local or private evaluation by legal teams.

Out-of-Scope Use

Do not use this candidate for:

unsupervised legal advice;
final court filings without attorney review;
citation generation without independent verification;
production deployment based only on this model card;
claims that layer removal improves legal capability without benchmark support.

Method Summary

The parent REAP-0.16 checkpoint was produced by routed-expert pruning from Qwen3.5-122B-A10B. This companion model then removes six decoder layers:

8, 9, 12, 13, 16, 17

Those layer indices were selected from a calibration-driven layer-importance pass. Public documentation intentionally does not identify private legal source items used for calibration or evaluation.

Evaluation Status

Status: research candidate.

The checkpoint is structurally complete and suitable for evaluation. Public aggregate legal benchmark tables are not yet attached to this card. Until those tables are published, treat the model as a candidate for research and internal evaluation rather than a validated production legal model.

Privacy Boundary

This public release intentionally excludes:

client names;
matter names;
legal source item titles;
private calibration prompts;
attorney-client communications;
privileged work product;
local hostnames, private paths, credentials, or tokens.

Limitations

Layer removal can produce different degradation patterns from expert pruning.
Strong fluency is not proof of legal reliability.
Public aggregate evaluation tables are still pending.
This is not an expert-merged checkpoint; it is a Layerdrop6 derivative of an expert-reduced REAP checkpoint.
All legal outputs require attorney review and source verification.

ProprietaryLegal/Qwen3.5-91b-a10b-LegalReap-Layerdrop6

Qwen3.5-91b-a10b-LegalReap-Layerdrop6