Skip to content
Other

Qwen3.8-27B-pi-AuxMix-v1

мердж с донорами и транспортными дельтами

1 min read

BF16-модель на 27,78 млрд параметров. Это мердж на базе Qwen/Qwen3.8-27B: subject Qwen3.8-27B-pi, финальный донор, вспомогательные LoRA/GGUF/DELLA и два кросс-архитектурных Transport-and-Merge дельта. Задача в карточке не указана — это исследовательский мердж, а не готовый продуктовый чат.

Ключевые особенности

  • База — Qwen/Qwen3.8-27B; итоговый вес 27,78B в BF16.
  • Основная формула: F = S + lambda_g * [D - kappa*(S-B)] + mu*C + nu*N; коэффициенты mu=0,75, nu=0,05, nu2=0,05.
  • Доноры: StartLuxLoRA 0,4, Winery 0,05, DavidAU 0,3, DellaMix 0,8, EfficientThink 0,5, DellaMix2 0,1.
  • N и N2 — кросс-архитектурные дельты от Nanbeige4.3-33B и SenseNova-Loop-9B-Preview, вмешаны с alpha=0,05.
  • Компоненты отбирали по forward NLL на 30 прокси-вопросах; часть коэффициентов изменена у 21 из 193 компонент.
  • Генерация, бенчмарки и safety-оценка не проводились; карточка прямо просит оценить модель перед использованием.

Пример

Готового кода запуска в карточке нет. Разумный сценарий — сравнить модель с subject на своих промптах и замерах NLL: карточка приводит только proxy-оценку на 30 вопросах, без примеров генерации.

ПодробнееProxy NLL: 0,0006 против 0,0023 у subject на 30 вопросах; улучшение 0,0016, но эти же вопросы использовали для выбора.
VRAM не указана; BF16-модель на 27,78B.
Ограничения: нет бенчмарков и safety-оценки; среди источников есть uncensored-файнтюны, отказы могут отличаться.
Ни одна изменённая строка весов не отличается от subject больше чем на 10% нормы строки.
Discuss Telegram 0 24

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 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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.

    # YFC-112358/Qwen3.8-27B-pi-AuxMix-v1
    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 YFC-112358/Qwen3.8-27B-pi-AuxMix-v1 --local-dir ~/models/Qwen3.8-27B-pi-AuxMix-v1
    # 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

New posts about versions of Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

YFC-112358/Qwen3.8-27B-pi-AuxMix-v1
Author
YFC-112358
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

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

Comments