---
title: Qwen3-16B-A3B-GGUF
canonical_url: "https://www.modelscope.cn/models/unsloth/Qwen3-16B-A3B-GGUF"
md_url: "https://www.modelscope.cn/models/unsloth/Qwen3-16B-A3B-GGUF.md"
repository: unsloth/Qwen3-16B-A3B-GGUF
last_updated: 2025-05-03
license: apache-2.0
model_type:
  - qwen3_moe
architectures:
  - Qwen3MoeForCausalLM
base_model:
  - kalomaze/Qwen3-16B-A3B
base_model_relation: quantized
library_name:
  - gguf
  - pytorch
frameworks:
  - pytorch
inference_backends:
  - "deploy_task text/emb"
  - "lmdeploy_turbomind 0.9.1"
  - "sglang 0.5.2"
  - "vllm 0.9.2"
downloads: 920
stars: 0
tags:
  - unsloth
  - gguf
---

# Qwen3-16B-A3B-GGUF

> Qwen3-16B-A3B-GGUF - unsloth 在 ModelScope 开源的模型。Qwen3-16B-A3B is a rendition of Qwen3-30B-A3B by kalomaze.

unsloth/Qwen3-16B-A3B-GGUF 是 ModelScope 魔搭社区上的机器学习模型，采用 apache-2.0 许可，基于 kalomaze/Qwen3-16B-A3B 构建，可用 deploy_task text/emb、lmdeploy_turbomind 0.9.1、sglang 0.5.2 部署。

- **Repository**: unsloth/Qwen3-16B-A3B-GGUF
- **License**: apache-2.0
- **Base model**: kalomaze/Qwen3-16B-A3B
- **Inference backends**: deploy_task text/emb, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2
- **Tags**: unsloth, gguf
- **Downloads**: 920
- **Stars**: 0
- **Last updated**: 2025-05-03

Source: https://www.modelscope.cn/models/unsloth/Qwen3-16B-A3B-GGUF

---

# Qwen3-16B-A3B

Qwen3-16B-A3B is a rendition of Qwen3-30B-A3B by [kalomaze](https://huggingface.co/kalomaze/Qwen3-16B-A3B).

A man-made horror beyond your comprehension.

But no, seriously, this is my experiment to:
- measure the probability that any given expert will activate (over my personal set of fairly diverse calibration data), per layer
- prune 64/128 of the least used experts per layer (with reordered router and indexing per layer)

It can still write semi-coherently without any additional training or distillation done on top of it from the original 30b MoE.
The .txt files with the original measurements are provided in the repo along with the exported weights.

Custom testing to measure the experts was done on a hacked version of vllm, and then I made a bespoke script to selectively export the weights according to the measurements.