Skip to content
Other

PixtralViT-Large-4

Tower и Projector из Mistral Large 4

1 min readupdated

Репозиторий immanuelpeter упаковывает визуальный энкодер Tower и Projector из Mistral Large 4 (mistralai/Mistral-Large-4-1T-A52B) для извлечения признаков изображений. Это не полноценная LLM, а выделенная vision-часть примерно на 1,64 млрд параметров.

Ключевые особенности

  • Tower: 30 блоков, hidden 2048, 16 голов по 128, SwiGLU 6144, RMSNorm с eps 1e-5, патчи 14 px, axial 2D RoPE, без class token и финальной нормы.
  • Projector сжимает патчи 2×2: RMSNorm(2048), merge до 8192, Linear(8192, 2048), Linear(2048, 6144), GELU и Linear(6144, 6144) без bias.
  • При 448 px на выходе 1024 Tower-токена шириной 2048, 256 merged-токенов шириной 2048 и 256 projected-токенов шириной 6144.
  • Все 276 тензоров сверены с родительским чекпоинтом через torch.equal; выходы Tower, merged и projected совпадают побитово в FP32 и BF16 при 448 и 1540 px.
  • Веса сохраняют Mistral Ecosystem Open-Weight License, Python-файлы — MIT.

Пример

Сценарий из карточки: изображение проходит через PixtralImageProcessor и PixtralVisionModel; при 448 px можно получить 256 merged-токенов шириной 2048 и 256 projected-токенов шириной 6144 для дальнейших задач извлечения признаков.

ПодробнееВ релизе 272 тензора vision-энкодера и 4 тензора projector.
Проверка включает 276 тензоров и сравнение выходов в FP32/BF16.
Это только Tower и Projector, без языковой модели.
Экспортный скрипт пиннит родительскую ревизию и переписывает имена тензоров в HF Pixtral.
Discuss Telegram 0 79

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 3.0 GB · ready files3.9 GB
  • FP8 / INT8recommendedweights 1.5 GB2.2 GB
  • 4-bitweights 0.9 GB1.4 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.

    # immanuelpeter/PixtralViT-Large-4
    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 immanuelpeter/PixtralViT-Large-4 --local-dir ~/models/PixtralViT-Large-4
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to PixtralViT-Large-4
  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 Mistral-Large-4-1T-A52B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 7All on Hugging Face

About the model

immanuelpeter/PixtralViT-Large-4
Author
immanuelpeter
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

790

Ask AI about this model

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

Comments