Skip to content
Computer vision

Kataguru-Sceptic-35B-A3B-Finnish-v3.2

финский MoE с переводами и STEM

1 min read

Kataguru-Sceptic-35B-A3B-Finnish-v3.2 — финская MoE-модель от kataguru на базе Kataguru Sceptic 35B-A3B v3.1. Она совмещает языковые, reasoning- и мультимодальные vision-возможности с полностраничным переводом EN ↔ FI; в архитектуре 256 экспертов на слой и 3B активных параметров на токен.

Ключевые особенности

  • 256 экспертов на каждый из 40 слоёв: 254 замороженных эксперта ядра #0–#253 и два новых финских #254–#255.
  • Эксперт #254 отвечает за финскую грамматику, морфологию, падежи и техническую терминологию; #255 — за литературу, нон-фикшн и двунаправленный перевод.
  • Ядро сохраняет reasoning, coding, вызов инструментов, PLC/CNC и 27-слойный vision encoder; контекст — 262 144 токена.
  • Нативный chat template: режимы /fi, /translate и управление глубиной reasoning через {REASON:…}.
  • Опубликованы полные BF16 master weights, а также NVFP4, W4A16-AWQ и GGUF APEX-квантизации.

Пример

Пример запроса: /translate Переведи страницу с английского на финский, сохраняя абзацы и технические термины. Для ответа на финском можно начать промпт с /fi.

ПодробнееБенчмарки базы: 15-Benchmark Mean 70.26%, GPQA Diamond 70.20%, BFCL V4 AST 83.67%.
BF16 требует ~72 GB VRAM; NVFP4/W4A16-AWQ — от 1x 32 GB, GGUF APEX-Medium (~13.6 GiB) — на 16 GB GPU.
Для строгой финской морфологии и вычитки карточка советует temperature=0.2–0.3 и presence_penalty=0.3.
Discuss Telegram 0 30

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 67 GB · ready files74 GB
  • FP8 / INT8recommendedweights 33 GB37 GB
  • 4-bitweights 19 GB21 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.

    # kataguru/Kataguru-Sceptic-35B-A3B-Finnish-v3.2
    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 kataguru/Kataguru-Sceptic-35B-A3B-Finnish-v3.2 --local-dir ~/models/Kataguru-Sceptic-35B-A3B-Finnish-v3.2
    # 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. Qwen3.5-35B-A3B-Basefamily root
  2. Qwen3.5-35B-A3B
  3. Kataguru-Sceptic-35B-A3B-v3.1
  4. Kataguru-Sceptic-35B-A3B-Finnish-v3.2

New posts about versions of Qwen3.5-35B-A3B-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 1All on Hugging Face
Merges 2All on Hugging Face

About the model

kataguru/Kataguru-Sceptic-35B-A3B-Finnish-v3.2
Author
kataguru
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

300

Ask AI about this model

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

Comments