Skip to content
Computer vision

AnuLM-Omni-400M

компактная any-to-any модель на семь задач

2 min read

AnuLM-Omni-400M от toonist — один декодер MoE на 456M параметров (232M активных на токен), обученный с нуля для хинди, английского и Python, а затем дообученный сразу на семи задачах. Рядом стоят два обученных проектора и два замороженных энкодера: Whisper-small для аудио и SigLIP для изображений. Модель рассчитана на одну потребительскую GPU и CPU.

Ключевые особенности

  • Один MoE-декодер: 20 слоёв, hidden 1024, 61 445 токенов, 456M параметров; 232M активных на токен.
  • Чат, ответы по найденным passages и маршрутизация сообщения в интент — в одном наборе весов.
  • Распознавание английской и хинди-речи: замороженный Whisper-small плюс аудиопроектор на 12,6M параметров.
  • Изображения: замороженный SigLIP, проектор и задачи captioning и VQA.
  • Заменяет пять отдельных копий декодера, которые раньше отвечали за разные навыки.
  • Лицензия CC BY-SA 4.0: коммерческое использование с атрибуцией; запуск на CPU и одной consumer GPU.

Пример

from anulm.inference.serve import Engine
chat = Engine("toonist/AnuLM-Omni-400M", "cuda")

from anulm.multimodal.vision_encoder import load_trained
vl, tower, tok = load_trained("toonist/AnuLM-Omni-400M", "cuda")
ПодробнееБенчмарки: WER en 7,9%, WER hi mean 41,9%, intent BANKING77 85,0, MASSIVE hi 80,33, VQA acc 12,6.
Веса: 456,4M параметров, bfloat16; запуск на одной consumer GPU и CPU.
Ограничения: 400M-модель часто ошибается в фактах — для фактов используйте reading по найденным passages; речь только en/hi, изображения — в основном стиль captioning и TextVQA.
Без safety tuning; обучение с нуля на hi/en/Python, затем совместный fine-tune на семи задачах.
Discuss Telegram 0 46

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.9 GB · ready files1.4 GB
  • FP8 / INT8recommendedweights 0.4 GB1.0 GB
  • 4-bitweights 0.2 GB0.8 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.

    # toonist/AnuLM-Omni-400M
    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 toonist/AnuLM-Omni-400M --local-dir ~/models/AnuLM-Omni-400M
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to AnuLM-Omni-400M
  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 AnuLM-Base-v2-400M (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 1All on Hugging Face

About the model

toonist/AnuLM-Omni-400M
Author
toonist
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

460

Ask AI about this model

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

Comments

Computer vision

UAV-PUR-8B

UAV-PUR-8B
Computer vision

JEV-27B-VL

JEV-27B-VL