David

logic65

11 1 3

huggingface666's profile picture

ShitheadPepperoni's profile picture

brrrock's profile picture

AI & ML interests

AI tinkerer in South Africa. I take language models apart on two 8GB gaming GPUs to see what's inside, then put smaller ones back together. Everything I break and everything I learn ships public. Support the tinkering: ko-fi.com/davida81328

Recent Activity

updated a model 41 minutes ago

logic65/Whittle-Qwen-3.8-35B-A3B-GGUF

updated a model about 21 hours ago

logic65/Whittle-Qwen-3.8-35B-A3B

upvoted an article 1 day ago

Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem

View all activity

Organizations

None yet

logic65 's collections 3

Whittle: research and dev

The compression ladder, recipes, adapters and unrepaired variants behind the family.

logic65/Qwen3.8-Whittle-dev

Updated Aug 19 • 1
logic65/Qwen3.8-Whittle-v40-16.8B

Text Generation • 18B • Updated Aug 18 • 64 • 1
logic65/Qwen3.8-Whittle-w50-18.3B-unrepaired

Text Generation • 18B • Updated Aug 18 • 106 • 1
logic65/Qwen3.8-p44w75-16.8B-unrepaired

Text Generation • 16B • Updated Aug 18 • 618 • 2

Qwen3.8 Whittle family

A 27B whittled on two 8GB GPUs: measured cuts, heals, and restorations. 48L = best untrained, Whittle-16B = best overall.

logic65/Qwen3.8-Whittle-16B

Text Generation • 16B • Updated Aug 18 • 2.18k • 4
logic65/Qwen3.8-Whittle-48L-cut

Text Generation • Updated Aug 18 • 1
logic65/Qwen3.8-p44w75-16.8B-unrepaired

Text Generation • 16B • Updated Aug 18 • 618 • 2
logic65/Qwen3.8-Whittle-v40-16.8B

Text Generation • 18B • Updated Aug 18 • 64 • 1

Whittle: models you can run

Compressed Qwen3.8-27B variants with weights published and tested. Start here.

logic65/Qwen3.8-Whittle-MoE-27B-A17.8B

27B • Updated 15 days ago • 24.4k • 122
logic65/Whittle-Next-26B-A3B

Text Generation • 27B • Updated 26 days ago • 581 • 3
logic65/Whittle-Next-27B-A3B

Text Generation • 28B • Updated 4 days ago • 5.38k • 25