logic65/Qwen3.8-Whittle-v40-16.8B

🤗 Hugging Face 来源text-generationapache-2.0激活 16.8B19 GBGGUF✓ 1 个校验和今天更新
一条命令提交

在你的模型文件夹旁边运行它。它会制作种子、将文件与 Hugging Face 比对,然后提交。你只需开始做种,并粘贴你账户中的密钥。它只读取你的文件,绝不修改。如果愿意,可以先阅读脚本。

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo logic65/Qwen3.8-Whittle-v40-16.8B ./model-folder
需要做种者 →

⚠️ 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-v40 (16.8B) best measured cut

🔬 RESEARCH PREVIEW. Smaller than every other cut in the family and

better than all of them. Chosen by measurement, not by assumption. No training.

40 layers of Qwen3.8-27B, kept in the exact [gdn, gdn, gdn, attention] pattern llama.cpp requires: 30 gated-DeltaNet layers, 10 full-attention stations, a true 3:1 ratio. Which layers survive was decided by a per-layer identity map (cosine between each layer's input and output residual stream over 80 probes): within every span between surviving attention stations, the three most load-bearing recurrent layers are kept and the rest are dropped.

Why this beats the earlier cuts

Measured with an 80-probe, 8-domain knowledge atlas, scored as the mean fraction of the intact model's next-token probability retained:

shape params layers retention
v40 (this) 16.8B 40 0.648
44L cut 19.2B 44 0.408
same size, attention crowded early 16.8B 40 0.379
same rule, 36 layers 15.1B 36 0.441

Two lessons are in that table. Choosing layers by measurement instead of removing contiguous bands is worth more than 2 billion parameters. And where the attention stations sit matters more than how many there are: three variants at identical size and identical selection rule score 0.648, 0.544 and 0.379 purely by station placement.

Against the intact 27B, this cut is at or above the original on agentic control (0.0922 vs 0.0885), Python (0.0030 vs 0.0023) and science (0.0035 vs 0.0034), holds 90% on math, and keeps 58% of multilingual recall where the 44L cut kept 3%. Its end-of-turn probability at a finished answer, the quantity that governs looping, stays at 0.46 and 0.44 versus the original's 0.53 and 0.33.

Honest state

No repair training has been done. It is a research preview: expect dulled confidence on hard recall and the family's known rough edges. C++ recall is the weakest domain, being distributed across layers no cut preserves. Anti-loop sampling is still recommended: --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.