Skip to content
Computer vision

dormice

VLM, который отвечает вероятностями

2 min readupdated

dormice — файнтюн Qwen/Qwen3.8-27B для вопросов с выбором ответа по изображениям. Модель возвращает вероятность для каждого варианта, а не генерирует текст. В репозитории — low-rank обновление весов (1.4 ГБ), 26-way linear head и движок Decision Index.

Ключевые особенности

  • Основа — Qwen/Qwen3.8-27B; vision tower заморожен, обучались LoRA rank-16 на линейных слоях языковой модели.
  • Ответ — выход 26-way head по hidden state последнего токена; softmax по вариантам при temperature 1, без декодирования букв.
  • До 12 вариантов A–L на запрос; изображения в нативном разрешении до 1 638 400 пикселей.
  • В full-attention слоях все токены промпта видят все токены, recurrent Gated DeltaNet остаются causal; обычное causal attention даёт другие ответы.
  • Нужен один GPU: bf16-веса занимают 54 ГБ, публичный прогон — около 58 ГБ.

Пример

from dormice import Engine
engine = Engine()
response, _ = engine(
    {},
    {"answer": {"type": "choice", "instructions": "Which object is closer?",
                "criteria": {"A": "the bicycle", "B": "the bench"}}},
    images=["street.jpg"])
print(response["answers"]["answer"])
ПодробнееПубличный прогон Decision Index vision 0.3: индекс 77.82 на десяти бенчмарках (11 410 из 11 410 вопросов); Winoground не запускался.
Сильные результаты: InfographicVQA 99.18, KIE 98.05, Mind2Web 95.17.
Слабые места: Hateful Memes (skill 54.82), R-Bench-M (29.97), MMMU-Pro vision (43.85) — вероятностям на таких вопросах лучше не доверять.
Загрузка с прогревом — около 109 с; медианная задержка 199 мс на RTX PRO 6000 (96 ГБ).
Это не заявка на SOTA: авторы прямо говорят, что разница с RSI-JEV-27b в пределах шума.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.3 GB · ready files1.9 GB
  • FP8 / INT8recommendedweights 0.7 GB1.2 GB
  • 4-bitweights 0.4 GB0.9 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.

    # plantwaterer/dormice
    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 plantwaterer/dormice --local-dir ~/models/dormice
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to dormice
  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

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

plantwaterer/dormice
Author
plantwaterer
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments