Modelpublic

0xSero/Qwen3.6-28B

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

34likes181downloads

Model Card

[!TIP] [Support this work →](https://donate.sybilsolutions.ai) · X · GitHub · REAP paper · Cerebras REAP

Qwen3.6-28B

At a glance


Base model	Qwen/Qwen3.6-35B-A3B
Format	BF16
Total params	28B
Active / token	—
Experts / layer	—
Layers	—
Hidden size	—
Context	—
On-disk size	56 GB

Which variant should I pick?

Variant	Format	Link
`Qwen3.6-28B` (this)	BF16	link
`Qwen3.6-28B-GGUF`	GGUF	link
`Qwen3.6-35B-GGUF`	GGUF	link

Please support my work: https://donate.sybilsolutions.ai

Qwen3.6-28B-REAP20-Opus-A3B

A 20%-expert-pruned + Opus-trace fine-tuned variant of Qwen/Qwen3.6-35B-A3B, produced via Cerebras REAP (Router-weighted Expert Activation Pruning, arXiv:2510.13999) followed by LoRA SFT on public Claude Opus reasoning traces.

Headline numbers

Metric	Base Qwen3.6-35B-A3B	This model (20% REAP + Opus SFT)	Δ
MMLU (200-sample lite)	`{{MMLU_BASE}}`	`{{MMLU_THIS}}`	`{{MMLU_DELTA}}`
GSM8K (100-sample lite)	`{{GSM_BASE}}`	`{{GSM_THIS}}`	`{{GSM_DELTA}}`
HumanEval (50 parse-rate)	`{{HE_BASE}}`	`{{HE_THIS}}`	`{{HE_DELTA}}`
Structured JSON parse (20)	`{{JSON_BASE}}`	`{{JSON_THIS}}`	`{{JSON_DELTA}}`
Mermaid render (10)	`{{MERM_BASE}}`	`{{MERM_THIS}}`	`{{MERM_DELTA}}`
AdvBench refusal (32)	`{{REFUSE_BASE}}`	`{{REFUSE_THIS}}`	`{{REFUSE_DELTA}}`

Architecture

—Base: Qwen3.6-35B-A3B (40 layers, 256 experts/layer, 8 routed + 1 shared active, qwen3_5_moe)
—After 20% REAP: 205 experts/layer kept, 51 experts/layer pruned → ~28B total params, still ~3B active
—Fine-tune: LoRA rank 32, α 64 on q,k,v,o,gate,up,down projections. bf16 weights after merge.

Pipeline

1.Calibration merge — 5,000 stratified samples from:
2./Users/sero/.../reap-expert-swap/dataset/calibration-20k.jsonl (general, coding, reasoning, etc.)
3.0xSero/structured-outputs-calibration-v1 (JSON / Mermaid / schema)
4.REAP observation (this fork's Qwen35Moe-aware observer, multi-GPU layerwise on 8× A100-40GB): {{OBSDURATION}}
5.REAP prune @ 20% using reap saliency metric, renormalized router weights, seed 42.
6.Opus-trace SFT via LLaMA-Factory + DeepSpeed ZeRO-3 (8× A100). LoRA 2 epochs on nohurry/Opus-4.6-Reasoning-3000x-filtered (2,326 reasoning trajectories with explicit <think>…</think>\nanswer structure).
7.GGUF — bf16, Q80, Q6K, Q5KM, Q4KM with imatrix from merged calibration.

Sidecar observations

REAP observation artifacts live in the separate dataset repo `0xSero/qwen3.6-35b-a3b-reap-observations`.

Known limitations

—Refusal behavior follows the base model plus Opus SFT; no explicit abliteration was applied in this release. The model will refuse straight adversarial probes at roughly base-model rates.
—Reasoning quality on GSM8K-style problems depends on the <think> chain-of-thought; short max-tokens limits hurt accuracy.
—Structured-output calibration is oversampled vs. base mix (JSON/Mermaid experts preferentially retained).

License

Apache 2.0, inherited from base model. This checkpoint is a derivative work; please preserve attribution.

License & citation

License inherited from the base model.

bibtex

@misc{lasby2025reap,
  title  = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author = {Mike Lasby and Ivan Lazarevich and Nish Sinnadurai and Sean Lie and Yani Ioannou and Vithursan Thangarasa},
  year   = {2025}, eprint = {2510.13999}, archivePrefix = {arXiv}
}