Skip to content
Computer vision

Gemma 4 31B IT

мультимодальный dense-флагман Google DeepMind

2 min readupdated

Инструктивно-донастроенная версия Gemma 4 31B: принимает текст и изображения, выдаёт текст. 60 слоёв, ~30.7B параметров, контекст до 256K токенов, словарь 262K, лицензия Apache 2.0. Вариант dense — MoE в семействе отдельный, 26B A4B.

Ключевые особенности

  • Понимание изображений: OCR (в том числе мультиязычный), разбор документов и PDF, графиков, интерфейсов, рукописного текста; переменные разрешение и соотношение сторон.
  • Режим размышления: включается через enable_thinking=True, вывод разбирается функцией parse_response.
  • Нативный function calling под агентные сценарии плюс генерация, дополнение и правка кода.
  • Гибридное внимание: локальное скользящее окно на 1024 токена чередуется с глобальным, в глобальных слоях — unified Keys/Values и p-RoPE.
  • Претрейн на 140+ языках, 35+ поддерживаются «из коробки»; нативная роль system.
  • Видео обрабатывается как последовательность кадров; аудио у 31B нет — только у E2B, E4B и 12B.

Пример

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "google/gemma-4-31B-it"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID, dtype="auto", device_map="auto"
)
ПодробнееБенчмарки IT-версии: MMLU Pro 85.2%, GPQA Diamond 84.3%, AIME 2026 без инструментов 89.2%, LiveCodeBench v6 80.0%, Codeforces ELO 2150, MMMU Pro 76.9%.
31B ориентирован на потребительские GPU и рабочие станции, но конкретных требований к VRAM карточка не приводит.
Аудиовход есть только у E2B, E4B и 12B; у 31B — текст, изображения и видео по кадрам.
Discuss Telegram 0 29

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 58 GB · ready files65 GB
  • FP8 / INT8recommendedweights 29 GB33 GB
  • 4-bitweights 16 GB19 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.

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.

    # Jamesl0l/gemma-4-31B-it
    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 Jamesl0l/gemma-4-31B-it --local-dir ~/models/gemma-4-31B-it
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gemma-4-31B-it
  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 gemma-4-31B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
Merges 3All on Hugging Face

About the model

Jamesl0l/gemma-4-31B-it
Author
Jamesl0l
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

290

Ask AI about this model

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

Comments