Skip to content
Computer vision

Qwen3.5-0.8B Heretic

расцензуренная версия модели, веса которой правили вручную

2 min readupdated

Это decensored-вариант Qwen3.5-0.8B-weight-space-programming от ML-Intern-lab, обработанный через Heretic v2.2.0.dev0 методом Arbitrary-Rank Ablation. Базовая модель примечательна тем, что её не обучали: frontier-LLM «Astra» прочитала веса Qwen3.5-0.8B, диагностировала патологию зацикливания и вручную отредактировала 131 072 веса (64 колонки MLP output в слоях 15 и 19, down_proj). Ни gradient descent, ни файнтюна — прямая правка. Размер — 0.85B параметров.

Ключевые особенности

  • Отказов 9/100 против 97/100 у исходной модели
  • KL-дивергенция всего 0.0744 — поведение сохранено
  • Правка воспроизводима бит-в-бит: каждый пересобранный тензор сверен по sha256
  • От публичного Qwen3.5-0.8B отличаются только два тензора — down_proj в слоях 15 и 19
  • MATH-500: 51.0% против 5.0% у базовой модели, обрезанных на 32k токенах ответов 31.6% вместо 95%

Пример

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

ckpt = "xtools-at/Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic"
tok = AutoTokenizer.from_pretrained(ckpt)
llm = LLM(model=ckpt, dtype="bfloat16", max_model_len=34816)
prompt = tok.apply_chat_template([{"role": "user", "content": "What is $2^{10} + 24$?"}], tokenize=False, add_generation_prompt=True, enable_thinking=True)
print(llm.generate([prompt], SamplingParams(temperature=0.6, top_p=0.95, max_tokens=32768))[0].outputs[0].text)
ПодробнееПараметры ARA: слои 5–23, neighbor_count 11, preserve_good_behavior_weight 0.8805.
Замеры: vLLM, thinking on, temperature 0.6, top_p 0.95, top_k 20, presence_penalty 1.5, до 32 768 новых токенов, seed 0.
Полный прогон MATH-500 (500 задач × 4 сэмпла) занял 51 минуту на NVIDIA H200.
Ограничение: это мультимодальный чекпоинт, но vision-башня текстовыми промптами не используется.
Discuss Telegram 0 68

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.6 GB · ready files2.2 GB
  • FP8 / INT8recommendedweights 0.8 GB1.4 GB
  • 4-bitweights 0.4 GB1.0 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # xtools-at/Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download xtools-at/Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic --local-dir ~/models/Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. Qwen3.5-0.8B-Basefamily root
  2. Qwen3.5-0.8B
  3. Qwen3.5-0.8B-weight-space-programming
  4. Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic

New posts about versions of Qwen3.5-0.8B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 9All on Hugging Face
Merges 1All on Hugging Face

About the model

xtools-at/Qwen3.5-0.8B-weight-space-programming-UNCENSORED-heretic
Author
xtools-at
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

680

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments