Downloads · 30 days
166
62% of all-time downloads
ProprietaryLegal/minimax-m3-legal-reap-0.25
minimax-m3-legal-reap-0.25 is a text generation model from ProprietaryLegal. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as other.
Published by PLI Labs, the open research arm of Proprietary Legal Intelligence. PLI builds AI systems around a legal organization's own work product, combining retrieval, staged workflows, validation, and attorney rev…
Downloads · 30 days
166
62% of all-time downloads
All-time downloads
267
Public
Parameters
323B
646 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors646 GB · 100%
How the weights are stored.
BF16323B · 100%
From the Hugging Face model README
<!-- pli:blurb:end --> <!-- pli:update:2026-10-27:start -->Published by PLI Labs, the open research arm of Proprietary Legal Intelligence. PLI builds AI systems around a legal organization's own work product, combining retrieval, staged workflows, validation, and attorney review. Visit proprietarylegal.ai to explore our work and discuss what your organization needs.
PLI Labs update 2026-10-27: our DeepSeek-V4.1-Flash on 2× DGX Spark work is now fully public — dsv41-offload (GitHub, every component with production, main and per-experiment branches), with focused repos for the measurement harness and sampled drafting. Status and lessons: 2026-10-27 post.
MiniMax-M3 Legal REAP 0.25 is a PLI Labs research artifact: a text-only, expert-pruned derivative of MiniMax-M3, produced to test whether a very large sparse mixture-of-experts model can be compressed around legal work (drafting, source-grounded synthesis, adversarial review, abstention, and long-context document reasoning) rather than around coding benchmarks.
This is an early, unhealed research checkpoint. It has not been recovery-tuned, it carries no published evaluation numbers, and it is held under review. It is not a legal assistant, and its outputs may be degraded relative to the base model. All legal use requires attorney supervision and independent source verification.
Built with MiniMax M3.
| Status | Research artifact — unhealed raw prune, under review |
| Evaluation | None published. No aggregate benchmark tables exist yet |
| Healing | Recovery fine-tune (heal) is the next pipeline stage — not yet applied |
| Verification | Held-out KL / faithfulness verification is pending |
| Recommended use | Research and internal evaluation only |
The public claim is deliberately narrow: this is a structurally complete, legal-calibrated expert-prune of MiniMax-M3, released as a research artifact so the method and the checkpoint can be inspected. It is not a benchmark-backed release and should not be treated as one until healing and verification complete and aggregate results are published.
| Field | Value |
|---|---|
| Model family | MiniMax-M3 (native multimodal MoE; this derivative is text-only) |
| Publisher | PLI Labs / Proprietary Legal Intelligence |
| Base model | MiniMaxAI/MiniMax-M3 |
| Pruning method | REAP-style routed-expert pruning (observation-driven saliency) |
| Prune ratio | 0.25 (floor-enforced, per-layer adaptive) |
| Working precision | bf16 |
| Format | Transformers / safetensors checkpoint (original split-expert layout) |
| Total parameters | ~323B after prune (base ~428B) |
| Active parameters | ~23B (top-4 routing preserved — active count essentially unchanged) |
| Intended domain | Lawyer-supervised legal drafting, review, and analysis |
| Contact | [email protected] |
The base MiniMax-M3 text backbone has 60 decoder layers: layers 0–2 are dense (no routed experts) and layers 3–59 are MoE, with 128 routed experts plus 1 always-on shared expert per MoE layer and top-4 sigmoid routing.
This derivative:
MiniMaxM3SparseForCausalLM text stack. MiniMax Sparse
Attention (MSA), the learned router, routing bias, layernorms, embeddings, and
the MTP modules are copied unchanged. The router weight and routing-bias rows for
dropped experts are removed by row-slicing (no numeric renormalization).This artifact is intended for:
The intended deployment pattern is local or on-premises evaluation on hardware controlled by the lawyer, firm, or legal organization.
Do not use this artifact for:
The prune direction is legal, not code. Routed-expert saliency was measured by running a private legal-domain calibration corpus through an observer that records, per expert, the REAP saliency signal (router gate value combined with activation magnitude over the active-token set). The calibration corpus spans legal drafting, summarization, adversarial review, abstention, and long-context trial-transcript roles (975 rows across 13 length-bucketed components); its contents are private and not distributed. Per-role observations were combined with a role-weighted merge that emphasizes legal drafting.
A per-layer adaptive schedule then set each layer's retained-expert count under capability floors (protecting under-covered layers and low-confidence carriers), and the lowest-saliency experts were dropped per layer. The checkpoint was written by a streaming split-to-split serializer that rewrites the on-disk shards directly and never loads the full model, touching roughly one shard buffer plus one tensor of peak memory. See the accompanying GitHub documentation for method detail.
The saliency criterion follows REAP (Lasby et al., "REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression", arXiv:2510.13999).
The checkpoint stores routed experts in split per-expert form with per-layer
nonuniform expert counts (84-108 of the original 128 per MoE layer). A small
modeling_minimax_m3_vl_split.py ships in the repo: it subclasses the native
transformers MiniMax-M3 implementation, replacing only the routed-experts module
with per-expert linear layers (memory-lazy sparse-attention masks included), and
is wired via auto_map — load with trust_remote_code=True:
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"ProprietaryLegal/minimax-m3-legal-reap-0.25",
trust_remote_code=True,
torch_dtype="bfloat16",
device_map="auto",
)
The split-linear expert layout is deliberately compatible with bitsandbytes
4-bit quantization (BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4")),
which quantizes every per-expert linear; the packed native layout cannot be
NF4-quantized. Forward equivalence of the split implementation against the
native packed implementation is test-pinned (bit-identical in bf16, identical
top-k routing).
No evaluation numbers are published. This checkpoint is the raw, unhealed prune. Field evidence and internal research indicate that MiniMax-M3's experts are near-orthogonal (little merge headroom) and that the honest cold-drop safe band for legal work is roughly 0.16–0.25, with the losses concentrated in knowledge, faithfulness, refusal, and long-context behavior rather than surface fluency — axes that coding-style benchmarks do not see. A recovery heal and held-out KL / faithfulness verification are the required next gates before any quality claim.
| Item | Value |
|---|---|
| Base model | MiniMaxAI/MiniMax-M3 |
| Base revision | MiniMaxAI/MiniMax-M3 (local bf16 conversion of the official release; no upstream revision pin recorded) |
| Prune ratio | 0.25 (floor-enforced, per-layer adaptive) |
| Selection metric | REAP saliency (canonical reap metric) |
| Seed | 42 |
| Cut-plan hash | 6c2e2d74e5cc594f5f32d28cbcfdd31e573c995d2b807c03a95cd180a1f2f8de |
| Retained experts/layer | 84–108 of 128 (per-layer adaptive) |
| Dense layers preserved | 0, 1, 2 |
| Shard total size | 645,832,896,384 bytes (~601.5 GiB, 121 shards) |
| Transformers baseline | 4.52.4 (base config); load path fuses split experts at load |
A machine-readable provenance sidecar (reap_prune_provenance.json) and an
expert-count census accompany the checkpoint and record the base index hash, the
observation hash, the full per-layer cut plan, and the cut-plan hash.
This release does not include client names, client files, closed-client materials, attorney-client communications, privileged work product, private prompts, calibration row contents, internal hostnames or IP addresses, local file paths, credentials, or tokens. The calibration corpus is described only generically.
This is a derivative work of MiniMax-M3 and remains subject to the MiniMax
Community License (license_name: minimax-community); the upstream LICENSE is
redistributed with this checkpoint. In summary, and without replacing the license
text:
Review the upstream license in full before use, especially for any commercial deployment.
PLI Labs. MiniMax-M3 Legal REAP 0.25 model card. Proprietary Legal Intelligence, 2026. Derived from MiniMaxAI/MiniMax-M3. Pruning method after Lasby et al., arXiv:2510.13999.
“Your proprietary work product is your proprietary intelligence.” PLI Labs is the open research arm of Proprietary Legal Intelligence. We publish legal model prunes, serving recipes for owned hardware, and open-source tools so readers can inspect our methods and test our claims before hiring us.
Founded by a practicing trial lawyer, PLI builds AI systems around a legal organization's own work product: retrieval, staged workflows, validation, and attorney review. The approach grew from systems running on live matters in a litigation practice. Our consulting work supports deployment through approved frontier APIs, private cloud, or open-weight models on hardware you own. Explore the research or contact us about your organization's work.
Website · PLI Labs · Research · GitHub · Hugging Face · Contact
Proprietary Legal Intelligence LLC is a technology consultancy, not a law firm, and does not provide legal services or legal advice.
<!-- pli:footer:end -->