Qwen3.5-0.8B-weight-space-programming-GGUF

Qwen3.5-0.8B Weight-Space Programming (ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming) is an Apache-2.0 open-weight multimodal checkpoint reconstructed by ML Intern from RikiyaT's Sparks of Weight Space Programming research. It demonstrates direct LLM-driven weight editing without gradient descent, training, or fine-tuning. Starting from Qwen/Qwen3.5-0.8B, the model's reasoning-loop pathology was addressed by directly modifying 131,072 weights across two MLP layers while preserving the vision tower. Evaluated on an NVIDIA H200 across all 500 MATH-500 problems with four generations per problem, it improves accuracy from 5.0% to 51.0% and reduces responses truncated at the 32,768-token limit from 95.0% to 31.6%. The results demonstrate substantial mathematical reasoning improvements through targeted weight-space editing alone.

[!NOTE] MTP — Speculative Decoding: No

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp