The obtained qwen1.5-1.8b-chat-rot_q4_0.mllm file size does not match the one in the ModelScope repository.

[singlecatlmx]

Hello,

I'm trying to reproduce the following two model files from the v1-backup:

• qwen1.5-1.8b-chat-rot_q4_0.mllm
• qwen1.5-1.8b-chat-rot_qnn.mllm

I strictly followed the instructions step by step in:

• QNN backend README
• Model conversion guide

However, the generated qwen1.5-1.8b-chat-rot_q4_0.mllm file is 1.0 GB in size, which does not match the file size of the official version hosted on ModelScope:
🔗 mllmTeam/Qwen1.5-1.8B-Chat on ModelScope

• My generated file: qwen1.5-1.8b-chat-rot_q4_0.mllm (1.0 GB)

• Official ModelScope file: qwen1.5-1.8b-chat-rot_q4_0.mllm (3.17GB)

Whether there are additional quantization or export steps I might have missed? How did you obtain the model files?
Thank you for your help!

[coderabbitai]

📝 CodeRabbit Plan Mode

Generate an implementation plan and prompts that you can use with your favorite coding agent.

• Create Plan

Examples

• Example 1
• Example 2

---

🔗 Similar Issues

Possible Duplicates

• 关于量化的疑问 #296

Related Issues

• Reason why mllm model file size is big #271
• 视觉权重无法找到，请问 demo_qwen2_vl_npu.cpp 中的模型文件如何获取 #535
• Tensor 'model.layers.0.self_attn.q_proj.bias' size mismatch. Code wants 8192, file has 2048 #507
• Where can I see the model structure of mllm file? #277

👤 Suggested Assignees

• kjh2159
• 970814
• oneseer
• yangyyj

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!

[chenghuaWang]

I think it's reasonable for a 1.8B model to use 1 GB, but 3.17 GB suggests something might be wrong, or perhaps some parameters weren't properly pruned/removed. @Sp0tless @oreomaker, we need to check it later.

[virus188]

我也发现了这个问题，并且自己导出的qwen1.5-1.8b-chat-rot_q4_0.mllm输出的是乱码

[chenghuaWang]

@virus188 你好。我们现在全面升级了 NPU 的后端，现在可以在 X86 的机器上 Convert 出来 Qnn 的模型了。您可以参考文档： https://ubiquitouslearning.github.io/mllm/qnn_backend/aot_execute.html 详细的技术报告在：https://chenghuawang.github.io/News/2026-01-29-mllm-qnn-aot-support/