Skip to content
Other

Qwen3-8B-5layer-MoE

a compact MoE drafter for faster generation

2 min read

This model is a specialized drafter for speculative decoding (a method for speeding up large language model generation), built on top of Qwen3-8B using SpecForge (DSpark). Its architecture features just 5 target layers, where each FFN (feed-forward network layer) is converted into a Mixture of Experts with 8 experts of width 24576 and two active experts per token. It has a total of 12.45B parameters with 3.39B active. The project is tailored for researchers and engineers working on LLM (large language model) inference speed optimization.

Key features

  • Architecture specifications: 5 target layers (1/9/17/25/33) with a projector and Markov head.
  • MoE structure: 8 experts without a shared expert, top-2 routing.
  • Parameters: 12.45B total and 3.39B active parameters per token.
  • Training: 1 epoch on the perfectblend dataset with an 8192 context and the qwen3-instruct template.

What you can do

  • Inference speedup: Use the model as a drafter for speculative decoding of Qwen3-8B.
  • Pipeline optimization: Deploy the model using a special inference library fork for MoE drafters.
  • Research: Study the behavior of 5-layer MoE architectures in text generation tasks.

Example

Running and serving the model requires a special package from the SpecForge repository fork (dev-moe-serving branch) with support for Qwen3MoeDSparkModel.

Read moreDetails
Base model: Qwen/Qwen3-8B
Architecture: 5 layers, MoE 8x24576, top-2
Training: 1 epoch, lr 6e-4, 9867 steps on GB300 nodes
Requirements: dev-moe-serving package for inference
Discuss Telegram 0

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 23 GB · ready files26 GB
  • FP8 / INT8recommendedweights 12 GB13 GB
  • 4-bitweights 6.5 GB7.7 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.

    # yushengsu/Qwen3-8B-5layer-MoE-8E2A-FFN24576-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 yushengsu/Qwen3-8B-5layer-MoE-8E2A-FFN24576-v1 --local-dir ~/models/Qwen3-8B-5layer-MoE-8E2A-FFN24576-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

  1. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. Qwen3-8B-5layer-MoE-8E2A-FFN24576-v1

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

yushengsu/Qwen3-8B-5layer-MoE-8E2A-FFN24576-v1
Author
yushengsu
Category
Other
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