jakeatx/slimder-qwen38-ream288-depth32-agentic

SLIMDER Qwen3.8 Depth-32 Agentic REAM-288

This is the BF16 Depth-32 structural winner compressed from 384 to 288 routed experts per layer with an agentic-weighted REAM calibration.

Lineage

• —Source: sjakek/slimder-qwen38-reap384-depth32-mb2-3-4-5
• —Pinned source revision: 2bb3209bb3c8424105985c5c971e7a2f22f1b61d
• —Structural depth: 32 transformer layers
• —Routed experts per layer: 288
• —Experts selected per token: 10
• —Precision: BF16

Calibration

The merge used 1,024 sequences of length 512, weighted toward agentic behavior:

• —tool calling: 384
• —executable/code tasks: 256
• —retrieval: 128
• —multi-step reasoning: 128
• —general language: 128

Calibration SHA-256: 7738db0c59401e2b6c4050aa1a4e9481a13c12419814e68a8adcf00fb06220d8.

Method

Experts were clustered and merged independently in every layer using REAM similarity with router outputs enabled, gated similarity, REAP saliency weighting, and sequential hidden-state propagation. The run kept 288 of 384 expert slots per layer and saved a resumable atomic checkpoint after every promoted layer.

Status and limitations

This is an experimental research checkpoint. Structural integrity and generation smoke tests are recorded in the linked public run repository. The model still requires broader agentic, code-execution, tool-use, retrieval, long-horizon, and safety evaluation before production use.

Run artifacts: sjakek/slimder-qwen38-agentic-ream-runs-20260901.

<!-- qwen38-perian-lineage:start -->

Qwen3.8 Perian project lineage

This repository is retained in the Qwen3.8 Perian checkpoints collection. Its exact position in the lineage is: Checkpoint after depth reduction from 48 to 32 layers and routed-expert width reduction from 384 to 288. It retains the full PLE table and predates the final QLoRA.

The final Qwen3.8 Perian GGUF release combines three reductions and one post-training stage:

• —depth: 48 to 32 transformer layers;
• —routed-expert width: 384 to 288 experts per layer;
• —PLE n-gram capacity: 320,001,446 to 160,000,768 rows (50%, about 25.60B parameters removed), using activation-aware bigram and frequency-ranked trigram selections validated on a document-disjoint 5M-token holdout;
• —rank-32 QLoRA on 12,558 normalized traces spanning math/STEM reasoning, coding/debugging, agentic tool use, retrieval, and general multi-step reasoning. The trace mixture draws from several frontier-model families, including Fable 5, GLM 5.2, Kimi K3, Claude Opus 4.7, Qwen3.8-Max, and GPT-5.6-Sol. The final merged milestone was trained through 9,336,692 supervised assistant tokens.

Earlier checkpoints in this collection do not inherit later stages merely by being listed beside them; the stage statement above is authoritative for this artifact. <!-- qwen38-perian-lineage:end -->