Skip to content
Computer vision

Qwen3.6-35B-A3B-REAP25

lightweight Qwen3.6 variant with pruned MoE experts

2 min readupdated

A modified language model by inference-optimization, built on Qwen/Qwen3.6-35B-A3B. It removes 25% of the original 256 experts across each of the 40 MoE (Mixture of Experts — an architecture where only specific parts of the network process each token) layers, leaving the 192 most important experts per layer. Pruning was performed using the REAP method (activation-based expert importance scoring) and the llmcompressor tool, with calibration done on the RazorCal dataset. The model targets researchers and developers looking to lower hardware requirements and speed up inference on a large model without retraining weights.

Key features

  • Reduced expert count: removes the 64 least important experts in each of the 40 layers (leaving 7,680 out of 10,240 in total).
  • Preserved weights: remaining weights (attention, vision module, embeddings, and shared experts) stay in the original BF16 (high-precision floating-point format) precision.
  • Unchanged MTP: the multi-token prediction (MTP) layer is left uncalibrated and untouched.
  • Automatic re-indexing: the model configuration (`config.json`) is updated with the new expert count, and expert indices are renumbered sequentially.

What you can do

  • Run in pipelines: integrate the model into projects via the transformers library to test text generation workloads.
  • Benchmark performance: measure generation speed and memory usage against the base Qwen3.6-35B-A3B model.
  • Run compressed inference: deploy the model in environments that support modified MoE layer configurations.

Example

from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
    "inference-optimization/Qwen3.6-35B-A3B-REAP25", torch_dtype="bfloat16"
)
processor = AutoProcessor.from_pretrained(
    "inference-optimization/Qwen3.6-35B-A3B-REAP25"
)
Read moreDetails
Base model: Qwen/Qwen3.6-35B-A3B
Parameter count: ~27.89B
Optimization method: REAP (25% expert pruning)
Calibration dataset: RazorCal (2,048 samples)
Library: transformers
Discuss Telegram 0 25

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 files58 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.

    # inference-optimization/Qwen3.6-35B-A3B-REAP25
    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 inference-optimization/Qwen3.6-35B-A3B-REAP25 --local-dir ~/models/Qwen3.6-35B-A3B-REAP25
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.6-35B-A3B-REAP25
  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 Qwen3.6-35B-A3B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 1All on Hugging Face
Merges 4All on Hugging Face

About the model

inference-optimization/Qwen3.6-35B-A3B-REAP25
Author
inference-optimization
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

250

Ask AI about this model

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

Comments