Skip to content
Computer vision

UniverseNet Qwen3.8

сжатые MoE-версии для 16 ГБ VRAM

2 min read

Набор из трёх чекпоинтов от wasnaga на базе nvidia/Qwen3.8-Flash-Next-NVFP4: 512 routed experts × 48 слоёв. Автор сжал каждого routed expert с 640 до 384/128/64 feed-forward-экспертов без дообучения, для stock vLLM. Каждая папка — полный чекпоинт: Workstation, Prosumer и Desktop16GB.

Ключевые особенности

  • Три издания: Workstation (384, 45.28 GiB), Prosumer (128, 19.97 GiB), Desktop16GB (64, 13.64 GiB) — GPU-веса.
  • Router, attention / Gated DeltaNet, hyper-connections, shared experts, embeddings и vision tower — бит-в-бит копии Teacher; меняется только ширина routed experts.
  • Сжатие занимает ~41.22 мин на одной NVIDIA GX10, 58 808 калибровочных токенов, без градиентных шагов.
  • Workstation сохраняет уровень Teacher на прогнанных наборах; Prosumer и Desktop держат tool calling и следование инструкциям, но теряют многошаговую арифметику и код.
  • NVFP4 нативно работает на GPU NVIDIA Blackwell; на Ada / Ampere / Hopper не тестировалось.

Пример

hf download wasnaga/UniverseNet-Qwen3.8-Flash-Next-NVFP4 --include "Desktop16GB/*" --local-dir ./qwen38
vllm serve ./qwen38/Desktop16GB --served-model-name universenet-qwen38 --trust-remote-code --quantization modelopt --kv-cache-dtype fp8_e4m3 --kv-cache-dtype-skip-layers 48 --mamba-ssm-cache-dtype bfloat16 --max-model-len 32768 --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3

Замените Desktop16GB на Prosumer или Workstation.

ПодробнееБенчмарки на 62 ручных промптах: Workstation — 96.77% overall и 12/12 tools; Desktop16GB — 43.55% overall, но 12/12 tools.
В каждой папке лежит неизменённый model-fp8-mtp-ple.safetensors на 53.7 ГБ (в основном embedding-таблица ~51 ГБ плюс ~3 ГБ MTP-head, не используется при MTP off), не входит в GPU-в��са — нужен запас диска и host memory.
Ограничения: 10–12 промптов на набор, широкий 95% CI; часть tool-промптов пересекается с калибровкой; цифры весов из логов vLLM.
Discuss Telegram 0

Will it run?

Only with RAM offload

Even 4-bit is larger than your VRAM. Some layers can stay in system RAM (llama.cpp, ComfyUI), but it will be several times slower.

  • FP8recommendedweights 50 GB · ready files56 GB
  • 4-bitweights 28 GB31 GB

Tip: go 8-bit.FP8 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: Won't fit any home GPU — renting is cheaper.

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.

    # wasnaga/UniverseNet-Qwen3.8-Flash-Next-NVFP4
    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 wasnaga/UniverseNet-Qwen3.8-Flash-Next-NVFP4 --local-dir ~/models/UniverseNet-Qwen3.8-Flash-Next-NVFP4
    # 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. Qwen3.8-Flash-Nextfamily root
  2. Qwen3.8-Flash-Next-NVFP4
  3. UniverseNet-Qwen3.8-Flash-Next-NVFP4

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

Quantised versions 12All on Hugging Face
Fine-tunes 8All on Hugging Face

About the model

wasnaga/UniverseNet-Qwen3.8-Flash-Next-NVFP4
Author
wasnaga
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments