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

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

Qwen3.5-0.8B with two weight matrices edited by an LLM, with no training, to stop its thinking loops and do better at math. An experimental model for researchers and small model users.

Made by RikiyaT on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is Qwen3.5-0.8B with its weights edited by an LLM that the card calls Astra, with no gradient descent and no training. Astra could run the model forward only 50 times, on practice math questions. It linked the base model's thinking loops to a single direction that the MLPs in layers 15 and 19 write out. It then removed that direction from 64 columns of each layer's down projection. That touches 131,072 weights, 0.016 percent of the model.

Who it suits: Researchers who study model editing, and people who want a tiny 0.8B model that finishes its math reasoning instead of looping.

Industry | Who uses it | How they use it | Example
Research | Model editing researchers | Study an edit made without training | Check the two changed tensors against the base
Software | Developers | Run a small model for math prompts | Serve it with vLLM
Education | Students | Study why thinking loops happen | Read the card's account of the edit

Skip it if you need broad, proven quality, because the card describes an experiment and reports math and instruction benchmarks only.

Where to use it:

• Python with vLLM, as in the card's usage example
• Python with the Transformers library, which the page lists
• A hosted demo on Hugging Face Spaces that compares the base model with the edited one

Use case | Who | What they get back
Math problem solving | Developers | Step by step answers that usually finish within the token limit
Edit verification | Researchers | Matching checksums for the two changed tensors
Small model testing | Hobbyists | A compact model to run on modest hardware

How to set it up:

• Python and the Hugging Face tools
• vLLM, if you follow the card's example
• About 1.7 GB of disk for the BF16 weights, as an estimate

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Load the model in vLLM with the settings from the card.

pip install -U huggingface_hub
hf download RikiyaT/Qwen3.5-0.8B-weight-space-programming

How to use it: Build the prompt with the tokenizer's chat template and thinking turned on, then sample with the card's settings. The card asks for step by step reasoning and a boxed final answer.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

ckpt = "RikiyaT/Qwen3.5-0.8B-weight-space-programming"
tok = AutoTokenizer.from_pretrained(ckpt)
llm = LLM(model=ckpt, dtype="bfloat16", max_model_len=34816)

sp = SamplingParams(n=1, temperature=0.6, top_p=0.95, top_k=20, presence_penalty=1.5,
                    max_tokens=32768, seed=0)
print(llm.generate([prompt], sp)[0].outputs[0].text)

Tool | What it does
LLM | Loads the model in vLLM
SamplingParams | Sets temperature, top p, top k and the token limit
apply_chat_template | Builds the prompt, with thinking enabled

Example: Ask it a short arithmetic problem and request step by step reasoning with a boxed final answer. The edited model should think through it and finish, where the base model tends to loop until the token limit.

For newcomers and experts: If you are new to this, try the hosted demo first. Enter a math problem in the demo and compare how the two models handle it.

If you know the field, only two tensors differ from the base, the down projection weights of layers 15 and 19, and the card gives checksums for them. The post evaluated a text only export, while this repository holds the full multimodal release with identical language model tensors.

Astra probes the base model's thinking loops
It finds one direction in two layers
It subtracts it from output columns
The edited model finishes more answers

Title: Math and instruction scores, base against edited
MATH-500 base: 7.9 percent
MATH-500 edited: 48.1 percent
GSM8K base: 13.8 percent
GSM8K edited: 67.1 percent
IFEval base: 21.0 percent
IFEval edited: 62.3 percent

Good to know:

• The share of answers that hit the token limit fell from 93.0 percent to 33.1 percent in the author's post.
• An independent re-run on full MATH-500 scored 51.0 percent, with 31.6 percent hitting the limit.
• The scores use thinking mode and up to 32,768 tokens, so shorter limits will differ.
• License: apache-2.0.
• Cost to run: none named, only your own hardware.
• Data it reaches: the prompts you give it. The source names no account or outside service.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 873M
• Weights: about 1.7 GB in BF16; about 437 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,374
• Likes: 137
• License: apache-2.0
• Based on: Qwen/Qwen3.5-0.8B
• Quantized versions: 5
• Fine-tunes: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, ml-intern, weight-space-programming, conversational

Links

• Model: https://huggingface.co/RikiyaT/Qwen3.5-0.8B-weight-space-programming
• Files: https://huggingface.co/RikiyaT/Qwen3.5-0.8B-weight-space-programming/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download RikiyaT/Qwen3.5-0.8B-weight-space-programming

Then follow the usage in its model card. Nothing in this kit runs by itself.