Skip to content
Language models

Keural Nova 1.1

корейский файнтюн Qwen MoE с честными компромиссами

2 min read

Keural Nova 1.1 — instruction-tuned модель от MKD на базе Qwen/Qwen3.6-35B-A3B (Apache 2.0): 35B параметров MoE и ~3B активных на токен. Это SFT через LoRA, слитую в веса, на 253 528 примерах (~85% корейский, 15% английский replay) для корейского диалога и понимания текста.

Ключевые особенности

  • Заточен под корейский диалог, понимание и генерацию; v1.1 — финальная точка после 2 эпох и 15 846 шагов.
  • База — MoE Qwen3.6-35B-A3B: 35B параметров, ~3B активных на токен; LoRA rank 32, α 64, слита в веса.
  • Адаптированы attention, Gated-DeltaNet, shared experts и все 256 routed experts в каждом слое; router и gates заморожены.
  • Не для генерации кода: HumanEval pass@1 падает с 62,20 до 11,59. Причина — повреждённые отступы в части обучающих примеров.
  • Наследует vision tower базы, но vision не обучали и не оценивали — считайте модель текстовой.
  • Контекст: SFT до 8 192 токенов; база до 262 144, через YaRN — до 1M. Подача до 1M проверена в проде, accuracy long-context не измерялась.

Пример

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "mkd-dev/Keural-Nova-1.1"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
messages = [{"role": "user", "content": "한국의 전통 발효 음식에 대해 설명해 주세요."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, temperature=0.7, top_p=0.95)
ПодробнееБенчмарки: KoBEST 67,29 против 62,88 у базы (+4,41), MMLU 82,37 против 83,83 (−1,46), HumanEval 11,59 против 62,20.
Обучение: LoRA rank 32, α 64, 2 эпохи, 15 846 шагов, effective batch 32, LR 1e-4 cosine, bf16, DeepSpeed ZeRO-2, 2×H200.
Ограничения: просадка на KMMLU/HAE-RAE/MMLU на ~1–2 пункта, сильная деградация кода; long-context accuracy без отдельного бенчмарка; vision не обучалась.
GSM8K 86,88 — рост в основном за счёт формата ответа, а не +55 к reasoning; в карточке об этом прямо предупреждают.
Discuss Telegram 1 254

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 files68 GB
  • FP8 / INT8recommendedweights 33 GB34 GB
  • 4-bitweights 18 GB20 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.

Language models: with a 8192-token context.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent RTX A6000

    $0.43 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    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.

    # mkd-dev/Keural-Nova-1.1 — RTX A6000, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve mkd-dev/Keural-Nova-1.1 --max-model-len 8192 --quantization fp8
    Add more LoRAs to Keural-Nova-1.1
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

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.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of Qwen3.6-35B-A3B (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 4All on Hugging Face

About the model

mkd-dev/Keural-Nova-1.1
Author
mkd-dev
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

2541

Ask AI about this model

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

Comments

Computer vision

squad-0.8b

squad-0.8b