Skip to content
Other

K-Decision-9B

корейский движок решений с калиброванными вероятностями

1 min read

K-Decision-9B v0.1 — первый публичный релиз от ONYULTEA: специализированная под корейский язык модель на базе Qwen/Qwen3.5-9B (8,95 млрд параметров). Это не чат-LLM: она принимает state и типизированные вопросы, а на выходе в одном forward pass даёт структурированное решение с вероятностями по всем вариантам.

Ключевые особенности

  • Не генерирует текст: предсказывает логиты символов ответа у маркеров в JSON-скелете, без цикла декодирования.
  • Поддерживает choice (2–26 вариантов) и noul (да/нет); softmax ограничен легальными символами, сумма вероятностей = 1.
  • Обучение и сервис — на одной 24GB consumer GPU: int8 LoRA, r=32; int8 выбран вместо 4-bit из-за точности вероятностей.
  • Специализация: корейские бытовые домены — администрация, медицина, CS, инфраструктура, образование, работа, агенты; языки — корейский и английский.
  • Веса Apache-2.0; есть сервер /v1/systemone, совместимый с System One / Jev.

Пример

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ONYULTEA/K-Decision-9B", torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("ONYULTEA/K-Decision-9B")
ПодробнееLocalLLaMA typed-decisions: 0.818 (fitted); KoBEST: 0.866; Korean holdout: 0.997.
Decision Index 0.3: 30.7 (raw 44.8), p50 168 ms, 0 ошибок на 140 тыс. запросов.
Лимиты: до 26 вариантов choice и 6 000 символов state; KMMLU ~30%; высокорисковые решения — только с человеком.
Веса Apache-2.0; обучение и сервис на одной 24GB GPU (int8 LoRA, r=32), ECE ~0.001 на holdout.
Discuss Telegram 1 37

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 17 GB · ready files19 GB
  • FP8 / INT8recommendedweights 8.3 GB9.7 GB
  • 4-bitweights 4.7 GB5.7 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 3080 · F16 — from RTX 4090

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.

    # ONYULTEA/K-Decision-9B
    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 ONYULTEA/K-Decision-9B --local-dir ~/models/K-Decision-9B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to K-Decision-9B
  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. Qwen3.5-9B-Basefamily root
  2. Qwen3.5-9B
  3. K-Decision-9B

New posts about versions of Qwen3.5-9B-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 6All on Hugging Face
Merges 1All on Hugging Face

About the model

ONYULTEA/K-Decision-9B
Author
ONYULTEA
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

371

Ask AI about this model

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

Comments