Skip to content
Language models

Qwen3-8B-24SST-Sparse-AdamW

Qwen3-8B, у которой оставили половину весов

2 min readupdated

Это разреженная версия Qwen3-8B: в каждой группе из четырёх весов остаются только два ненулевых. Модель не просто обрезали — её дообучали методом 24SST dynamic sparse training на OpenWebText, чтобы оставшиеся веса адаптировались. Архитектура та же: 36 слоёв, hidden size 4096, 32 attention heads и 8 KV heads. На выходе — генератор текста на базе Qwen3-8B, который на GPU с sparse tensor cores (NVIDIA Ampere и новее) может работать быстрее и занимать меньше места. Пригодится тем, кто запускает Qwen3-8B локально и хочет проверить 2:4-разреженность на своём железе.

Ключевые особенности

  • 2:4 semi-structured sparsity: не больше двух ненулевых весов на каждые четыре подряд.
  • На железе со sparse tensor cores — примерно в 2 раза быстрее matrix multiplies за счёт пропуска нулей.
  • Примерно в 1,8 раза меньше хранение весов: 2 значения плюс 2 компактных 2-битных индекса вместо 4 полных значений.
  • Не one-shot pruning: разреженную сеть переобучали, чтобы сохранить больше качества плотной модели.
  • База — Qwen/Qwen3-8B, 36 слоёв, hidden size 4096, 32 attention heads, 8 KV heads.

Что можно сделать

  • Запустить генерацию текста через transformers по примеру из карточки и проверить скорость на Ampere или новее.
  • Использовать как замену Qwen3-8B в экспериментах, где важны память и скорость, если ваш рантайм умеет 2:4 sparsity.
  • Сравнить ответы с плотной Qwen3-8B на своих промптах: в карточке нет бенчмарков качества.
  • Изучить поведение 24SST dynamic sparse training на практике — модель обучена на OpenWebText с оптимизатором AdamW.

Пример

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "SparseMindAI/qwen3-8b-24sst-sparse-adamw",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("SparseMindAI/qwen3-8b-24sst-sparse-adamw")
ПодробнееБаза: Qwen/Qwen3-8B, 36 слоёв, hidden size 4096, 32 attention heads, 8 KV heads.
Разреженность: 2:4 semi-structured; обучение — CHTsnm dynamic sparse training (24SST) на OpenWebText, оптимизатор AdamW.
Заявлено ускорение matrix multiplies ~2× и хранение весов ~1,8× меньше на подходящем железе.
Бенчмарков качества и требований к VRAM в карточке нет.
Discuss Telegram 0 46

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 15 GB · ready files17 GB
  • FP8 / INT8recommendedweights 7.6 GB9.4 GB
  • 4-bitweights 4.3 GB6.0 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 · BF16 — from RTX 4090

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

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 A30

    $0.26 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.

    # SparseMindAI/qwen3-8b-24sst-sparse-adamw — A30, 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 SparseMindAI/qwen3-8b-24sst-sparse-adamw --max-model-len 8192 --quantization fp8
  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

  1. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. qwen3-8b-24sst-sparse-adamw

New posts about versions of Qwen3-8B-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 4All on Hugging Face

About the model

SparseMindAI/qwen3-8b-24sst-sparse-adamw
Author
SparseMindAI
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

460

Ask AI about this model

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

Comments