Skip to content
Computer vision

Qwen 3.8 Flash-Next

MoE-пакет, который влезает в 16 ГБ VRAM

2 min readupdated

TinoBruno выложил оптимизированные бинарные веса, манифест, мультимодальный vision-проектор (mmproj) и таблицу Per-Layer Embedding N-gram для Qwen 3.8 Flash-Next (GSQ-RCO-IQ1_M). Всё это отформатировано под MinnieTheMoEcher — bare-metal рантайм для агентных задач. Архитектура qwen4exp: гибрид SSM и attention со разреженным MoE.

Ключевые особенности

  • 48 слоёв: 36 слоёв Hybrid Gated DeltaNet SSM + 12 слоёв Grouped-Query Attention, Hyper-Connections (hc_mult = 4, D = 10 240) и PLE-таблица на 320M n-gram.
  • MoE: 256 routed экспертов на слой, top_k = 10 плюс 1 gated shared expert — всего 12 288 экспертов.
  • resident.bin (~4.15 ГБ) с dense-частью, shared-экспертами и output head целиком живёт в VRAM.
  • experts.bin (~23.42 ГБ) стримится из pinned host DRAM через асинхронный double-buffered DMA; PLE-таблица (~26.82 ГБ) подключается zero-copy из памяти хоста.
  • mmproj на BF16 (~865.5 МБ) добавляет нативное мультимодальное зрение.
  • Заявлено ~113 tok/s prefill и ~20.5 tok/s decode на 16 ГБ VRAM без нагрузки на CPU.

Пример

# Скачать модель через загрузчик MinnieTheMoEcher
python scripts/download_model.py --model qwen_flash_next

# Запустить рантайм
./moecher -m models/qwen3.8_flash_next/moecher_manifest.json --port 8001
ПодробнееСмешанная RCO-квантовка: от IQ1_M до Q6_K и BF16 по разным тензорам.
Железо из карточки: GeForce RTX 5060 Ti 16 ГБ + 32/64 ГБ DDR5.
Токенизатор BPE на 248 320 токенов, языки — английский и китайский.
Важно: веса рассчитаны на рантайм MinnieTheMoEcher, а не на transformers или vLLM.
Discuss Telegram 0 45

Will it run?

Runs in FP8 / INT8

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 27 GB · ready files30 GB
  • GGUF IQ1_Mweights 27 GB · ready files30 GB
  • FP8 / INT8recommendedweights 13 GB15 GB
  • 4-bitweights 7.5 GB8.8 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 5080 · BF16 — from RTX 5090

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.

    # TinoBruno/moecher-qwen-3.8-flash-next
    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 TinoBruno/moecher-qwen-3.8-flash-next --local-dir ~/models/moecher-qwen-3.8-flash-next
    # 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 moecher-qwen-3.8-flash-next (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

TinoBruno/moecher-qwen-3.8-flash-next
Author
TinoBruno
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

450

Ask AI about this model

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

Comments