Qwen3.5-104b-a10b-LegalReap is PLI Labs' conservative legal-domain expert-reduced Qwen3.5 candidate. It starts from Qwen3.5-122B-A10B and removes the stable low-saliency routed-expert core observed under a legal calibration workload, reducing each MoE layer from 256 routed experts to 216 while preserving the original 48-layer depth, hidden size, tokenizer, and top-8 routing pattern.

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

Release Positioning

PLI Labs builds legal AI systems for lawyers who need private, inspectable, and locally evaluable model workflows. This release applies that thesis to a large open MoE model: instead of pruning around coding benchmarks, the calibration direction was legal drafting, factual synthesis, source-grounded review, workflow-sensitive reasoning, and conservative refusal behavior.

The public claim is deliberately specific:

this is a legal-domain REAP candidate;
it is expert-reduced, not expert-merged;
the released files are suitable for research and controlled evaluation;
Qwen3.5-122B-A10B appears prune-resistant, so the cut is intentionally conservative;
public aggregate legal benchmark tables should be attached before broader validated-production claims are made.

Model Details

Field	Value
Model family	Qwen3.5 MoE
Publisher	PLI Labs / Proprietary Legal Intelligence
Release status	Research candidate
Base model	Qwen/Qwen3.5-122B-A10B
Base license	Apache-2.0
Format	Transformers/safetensors checkpoint
Architecture	Qwen3_5MoeForCausalLM
Pruning method	REAP-style routed-expert pruning
Router handling	Renormalized over surviving routed experts
Hidden size	3072
Layers	48
Routed experts	216 per 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	31,839
Indexed tensor bytes	207,972,470,784
Estimated total parameters	about 104B
Contact	nick@proprietarylegal.com

Intended Use

This candidate is intended for:

lawyer-supervised legal drafting and revision assistance;
long-form legal and factual synthesis;
source-grounded summarization and review;
retrieval-sensitive analysis;
legal workflow/tool-use evaluation;
conservative refusal or escalation when the record is incomplete;
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;
client-specific work where the user cannot verify the source record;
production deployment based only on this model card;
claims of general coding, chat, or benchmark superiority.

Method Summary

The calibration direction was legal work, not generic chat or coding. Public documentation intentionally does not identify private legal source items used for calibration or evaluation.

The pruning procedure:

Observe routed-expert usage under a legal calibration workload.
Rank routed experts by REAP-style saliency.
Select the stable low-saliency core rather than forcing a larger prune.
Remove 40 routed experts per MoE layer, reducing 256 experts to 216.
Renormalize routing over survivors.
Save a complete Transformers/safetensors checkpoint.

The main technical finding is as important as the artifact: Qwen3.5-122B-A10B appears prune-resistant. PLI Labs therefore treats 0.16 as the conservative expert-reduction lane and uses quantization, evaluation, and possible future expert-merge research as the next levers.

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.

The raw run configuration included machine-local paths and is not uploaded to this repository. A public-safe settings summary is maintained in the PLI Labs GitHub documentation.

Limitations

Strong fluency is not proof of legal reliability.
Expert pruning can preserve surface quality while degrading source fidelity or tool judgment.
Public aggregate evaluation tables are still pending.
This is not an expert-merged checkpoint; it is an expert-reduced REAP checkpoint.
All legal outputs require attorney review and source verification.

ProprietaryLegal/Qwen3.5-104b-a10b-LegalReap

Qwen3.5-104b-a10b-LegalReap