You can deploy this model and start earning money today!

logic65/
Qwen3.8-Whittle-48L-cut

Name: Qwen3.8-Whittle-48L-cut
Availability: InStock

Views

No views yet

⚠️ NEEDS POST TRAINING. A research preview, published as the record of a

compression method and its measurements rather than as a finished assistant. Further healing and instruction tuning runs are planned over time, and these checkpoints will improve as those land.

Qwen3.8-Whittle-48L-cut (20.8B) 🏆 best untrained cut

🔬 RESEARCH PREVIEW. The strongest zero-training variant in the Whittle

family: 16 of 64 layers removed by measured block pricing, no width pruning, no repair training, and nothing measurably broke.

The 48-layer cut of Qwen3.8-27B (blocks 4-11 and 24-31 removed, GGUF interval pattern preserved). Battery: 33/39 at 5 tok/s on two consumer 8GB GPUs. Code, arithmetic, commonsense and factual recall all hold; several factual probes score sharper than the intact model. This is the "smallest undamaged" point on the measured compression ladder, and the recommended starting base for further training experiments.

Agent-grade characterization (q8 GGUF, 16k ctx)

Full rerun on a second rig (2x RTX 3060): fencing 8/8, loops 0/6 with DRY disabled, battery 35/39 (33/39 on the first rig; the suite has roughly a 2-point noise floor). Zero loop generations including the greedy watchmaker bait puts this cut ahead of every trained variant on loop behavior. Notable: this cut does not contain blocks 8-11, yet stays loop-free in single-turn use, so loop resistance tracks total depth (48 layers, 12 of 16 full-attention stations) rather than any single block.

Family scoreboard (39-prompt greedy battery):

variant	params	battery	speed
48L cut (this)	20.8B	33/39	5 t/s
44L cut	19.2B	28/39	9-10 t/s
16.8B cut+shrink, unrepaired	16.8B	25/39	20.5 t/s
Whittle-16B healed	16.8B	36/39	18.5-20.9 t/s

Serve with anti-loop sampling: --dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 4 --repeat-penalty 1.15

Support this work

Independent research on consumer hardware. Every donation becomes A100 hours, and every A100 hour ends up as a public model or a public measurement. ☕ ko-fi.com/davida81328

Acknowledgements

Base model by the Qwen team (Apache 2.0). Cut and measured by David Aylward with Claude (Fable 5, Anthropic) as co-author.

logic65/
Qwen3.8-Whittle-48L-cut

⚠️ NEEDS POST TRAINING. A research preview, published as the record of a

Qwen3.8-Whittle-48L-cut (20.8B) 🏆 best untrained cut

🔬 RESEARCH PREVIEW. The strongest zero-training variant in the Whittle

Agent-grade characterization (q8 GGUF, 16k ctx)

Support this work

Acknowledgements

logic65/Qwen3.8-Whittle-48L-cut

⚠️ NEEDS POST TRAINING. A research preview, published as the record of a

Qwen3.8-Whittle-48L-cut (20.8B) 🏆 best untrained cut

🔬 RESEARCH PREVIEW. The strongest zero-training variant in the Whittle

Agent-grade characterization (q8 GGUF, 16k ctx)

Support this work

Acknowledgements

logic65/
Qwen3.8-Whittle-48L-cut