---

license: apache-2.0

base_model: jakeatx/slimder-qwen38-ream288-depth32-agentic-ngram50-compact

pipeline_tag: text-generation

tags:

• gguf
• qwen
• mixture-of-experts
• ream
• lora
• model-compression
• agentic

---

Qwen3.8 Perian GGUF

Qwen3.8 Perian is a deliberately compressed and post-trained Qwen3.8-Flash-Next

lineage. It combines depth pruning, routed-expert width reduction, and a

document-disjoint, activation-informed reduction of the PLE n-gram table, then

adds a rank-32 QLoRA trained on frontier-model reasoning trajectories.

What was reduced

• Depth: 48 transformer layers were reduced to 32 (16 layers removed;

33.3% fewer layers).

• Routed-expert width: 384 experts per layer were merged to 288 (96 fewer

expert slots per layer; 25% reduction), while retaining 10 selected experts

per token. The merge used layer-local router, similarity, and activation

saliency evidence from 1,024 calibration sequences of 512 tokens, weighted

toward tool use, executable code, retrieval, and multi-step reasoning.

• PLE n-gram capacity: 320,001,446 rows were reduced to 160,000,768 rows

(50%). This removed about 25.60B parameters. The frozen selection combined

activation-aware bigram heads 0-7 with frequency-ranked trigram heads 8-15

and was validated on a document-disjoint 5M-token holdout.

The resulting compact BF16 base contains 74,615,655,680 parameters and

150,511,416,232 tensor bytes before GGUF quantization.

QLoRA post-training

The compact model was post-trained with rank-32 QLoRA on **12,558 normalized

traces** covering coding and debugging, mathematics and STEM reasoning,

agentic/tool-use trajectories, retrieval, and general multi-step reasoning.

The mixture contains traces produced by several frontier model families,

including Fable 5, GLM 5.2, Kimi K3, Claude Opus 4.7, Qwen3.8-Max, and

GPT-5.6-Sol. Assistant reasoning, answers, tool calls, and end-of-message tokens

were supervised; prompts, tool definitions/results, and repeated context were

masked. The merged checkpoint used here is the rank-32 milestone at **9,336,692

supervised assistant tokens**.

The private normalized training corpus has 25,811,891 canonical rendered tokens

and 10,299,662 canonical supervised tokens. It remains private because its

upstream records do not share one uniform license or blanket downstream-use

clearance. The adapter source is jakeatx/ream-288-rank32-3m-adapter at revision

41947bd12f74482906995f47b5c8a7c5703cfca1, path

milestones/checkpoint-wall-9336692/adapter.

Files

| File | Size | SHA-256 | Recipe |

|---|---:|---|---|

| Qwen3.8-Perian-UD-Q4-K-XS.gguf | 44,842,893,088 bytes (44.84 GB) | 70896b81fe0c4c24757a6df4344ec28c133a8c1893573227c1fe5593fcc89058 | Unsloth UD-Q3_K_XL tensor recipe with every IQ2/IQ3 XS-family assignment promoted to IQ4_XS. |

| Qwen3.8-Perian-Q4-K-M.gguf | 51,589,880,608 bytes (51.59 GB) | 9e40972ffa1248543616dc915555dca8f412100135507b43d8dc95d97bf48def | Standard llama.cpp control. |

| Qwen3.8-Perian-UD-Q5-K-XL.gguf | 69,482,800,928 bytes (69.48 GB) | 9d22ba116ad68b3d2f018247c7299cc7b8e1a386b79a5de7e093d4dfc1f6c6d6 | Current Unsloth Qwen3.8 tensor assignments for retained blocks 0-31. |

| Qwen3.8-Perian-Q8-0.gguf | 80,656,975,648 bytes (80.66 GB) | e71c7e0e840bb4fe6e4d5684a412ee9e9c5d7cb029640ebe4555219e50e34d52 | Standard high-fidelity control. |

The GGUF converter and quantizer use a pinned, patched qwen4exp llama.cpp build

that preserves the compact PLE remap tensor as integer metadata. A compatible

runtime must understand the qwen4exp architecture and compact PLE remap.

Status and limitations

The GGUFs do not yet include a validated MTP draft head. Treat MTP attachment

as a separate runtime-compatibility milestone rather than assuming the draft

head from the unpruned 48-layer model is compatible with this 32-layer trunk.

The pruning, holdout, merge, and training gates passed, but these artifacts

should still be evaluated on the target runtime and benchmark suite before a

production deployment decision.