128 likesNew · 8d old
Tiny Qwen3.5 model gets math boost via weight-space edit
What is Qwen3.5-0.8B-weight-space-programming?
Researchers let an LLM directly tweak 0.016% of the 0.8 B-parameter Qwen3.5 weights, stripping out a looping pattern that made the model hit its token limit on math questions. The edited checkpoint lifts MATH-500 accuracy from under 8% to roughly 50% and cuts token-limit failures from 93% to about 33%.
Who it is for
Developers of low-resource math assistants can grab this checkpoint for a cheap, faster model that works without any additional training.
How to try it
Open the live demo at https://huggingface.co/spaces/akhaliq/weight-space-programming-demo or pull the model from Hugging Face.
The story so far
Early Signal first spotted RikiyaT/Qwen3.5-0.8B-weight-space-programming on November 8, 2026, with 128 likes. On its busiest day it gained 18 likes. The project is 8 days old. Early Signal files it under LLMs, Image.
Questions about Qwen3.5-0.8B-weight-space-programming
- How do I try Qwen3.5-0.8B-weight-space-programming?
- Open the live demo at https://huggingface.co/spaces/akhaliq/weight-space-programming-demo or pull the model from Hugging Face.