README.md

---
license: other
license_name: minimax-non-commercial
license_link: LICENSE
base_model: MiniMaxAI/MiniMax-M2.7
tags:
- moe
- expert-pruning
- reap
- legal
- llama.cpp
- gguf
- volta
language:
- en
---

# MiniMax-2.7-LegalReapV2Adaptive-BASE

**An adaptively expert-pruned, legal-domain-calibrated derivative of MiniMax-M2.7 — 35%
of experts removed with per-layer precision, servable on stock llama.cpp, including
pre-Ampere (Volta) hardware.**

This is the **BASE** (pre-recovery) release of the LegalReap V2 line: the structural
prune itself, published with full provenance before any post-training. Healed and
optimized variants build on this artifact.

## What makes this release interesting

- **Adaptive, not uniform.** Most published MoE prunes cut every layer by the same
  ratio. This model's per-layer retained-expert counts (116–198 of the original 256)
  were chosen from **measured expert saliency** on real legal workloads — layers that
  concentrate importance keep more experts; redundant layers give up more. Overall:
  **10,317 of 15,872 experts retained (65.0%)** across 62 MoE layers, top-8 routing
  preserved.
- **Calibrated on real professional work.** Saliency observation ran over a curated,
  16-bucket calibration corpus drawn from a practicing attorney's private legal
  workload — drafting, document analysis, financial tracing, RAG planning/judging,
  routing, summarization, and abstention roles across multiple context lengths. (The
  corpus itself is private and is **not** part of this release; no client material is
  embedded in, or recoverable from, a structural prune.)
- **A novel GGUF export path for adaptive MoE.** Non-uniform expert counts don't fit
  GGUF's uniform stacked-expert format, so this release ships via a **pad-to-uniform
  export**: each layer is padded to 198 experts with never-selectable dummy experts
  (finite large-negative selection bias), and the true per-layer retained counts are
  recorded in the GGUF metadata. The result loads and serves on **stock llama.cpp** —
  no fork required. The conversion patch (against upstream `convert_hf_to_gguf`) is
  included in this repo for reproducibility.
- **Runs where big MoEs usually don't.** The Q8_0 GGUF (189 GB, 809 tensors) serves
  comfortably on a 10× Tesla V100-32GB (sm_70) node with layer split — hardware a
  generation older than what most 200B-class MoEs assume — and is proportionally
  lighter everywhere else: **~22% smaller** than the equivalent unpruned M2.7
  quantization at every precision.

## Files

| File | What it is |
|---|---|
| `*.safetensors` + index | The pruned model, bf16, HF Transformers format (adaptive metadata in `config.json`) |
| `gguf/MiniMax-2.7-LegalReapV2Adaptive-BASE.Q8_0.gguf` | Q8_0 quantization, stock-llama.cpp servable |
| `llama-cpp-minimax-adaptive-gguf-export.patch` | Conversion patch (pad-to-uniform adaptive export) |
| `artifact_integrity.json` | Machine-checked integrity manifest (metadata consistency, shard coverage, tokenizer equivalence, tensor counts) |
| `LICENSE` | MiniMax non-commercial license (inherited from the base model) |

`config.json` carries the full per-layer map in
`reap_adaptive_retained_experts_by_layer` (62 entries, 116–198), with
`reap_adaptive_nonuniform_experts: true`.

## Method in brief

Expert saliency was observed layerwise (REAP-style weighted activation statistics)
over the legal calibration corpus on V100 hardware, per-layer retention ratios were
derived from the measured saliency distribution rather than a global constant, and
the prune was executed with router renormalization (seed 42, target ratio 0.35).
Integrity gates run at every step — metadata completeness, shard coverage, semantic
tokenizer equivalence against the base model, and tensor-count verification of both
exports (809/809).

## Status and roadmap

This is a **research preview** and the deliberate starting point of a documented
recovery pipeline: diagnostics battery → targeted recovery (router recalibration,
router-KD, adapter healing) → verified, evaluated releases. Formal held-out
evaluations accompany the healed variants rather than this structural BASE — that
ordering is intentional, so every published number lands on a clean, reproducible
footing.

## Provenance

- Base model: MiniMax-M2.7 (bf16), © 2026 MiniMax, used under its non-commercial license.
- F16 GGUF sha256: `3994ce00d215e1aee63056e64e2e54ff07b3cb0a82fc71f5d82a417c79e5892c`
- Q8_0 GGUF sha256: `2d89f5dd56a5198276d5c5d947b59ca9d94ba586f856cf7884e2ab6e76b99da3`
- Built 2026-07-07 by Proprietary Legal Intelligence (PLI Labs).

## License

Inherits the MiniMax **non-commercial** license of the base model (see `LICENSE`):
non-commercial use, modification, and redistribution permitted with notice;
commercial use requires prior written authorization from MiniMax.