windows 11 intel npu - nexa sdk BUG

[paoIpao]

Pre-checks

• I searched existing issues
• I’m using the latest NexaSDK release

What happened?

I install nexa sdk.
when I execute (or any other models): nexa infer NexaAI/llama3.2-3B-intel-npu
I write something and works and response.
Then closed and when reopen the windows and execute: nexa infer NexaAI/llama3.2-3B-intel-npu
then I write something and the response is seems blank (or some character that I not see).

Then if close, wait some time and re-open window then it's start work againt.

If I execute: nexa infer NexaAI/deepSeek-r1-distill-qwen-7B-intel-npu
the problem on first time that I write something start report some random characters and then infinite "cryptocryptocrypto"...
Hen if close and reopen and write something the response is seems blanks (or some characters that I not see).

Steps to reproduce

1.Install nexa sdk on windows 11 - intel npu
2.execute: nexa infer NexaAI/llama3.2-3B-intel-npu (or nexa infer NexaAI/deepSeek-r1-distill-qwen-7B-intel-npu)
3. write something and see response
4. bug/problem on response with invalid characters/problem decode or other similar.

Logs & stack traces

Model(s) and quantization

deepSeek-r1-distill-qwen-7B-intel-npu

NexaSDK version

NexaSDK Bridge Version: v1.0.38-rc5 NexaSDK CLI Version: v0.2.65

Install method

from source

OS and version

Windows 11

Hardware / accelerator

NPU - Yoga Slim 7 14ILL10 - Type 83JX

Additional context

[mengshengwu]

Hi, thanks for reporting this issue.

The behavior you describe is caused by model quantization. Currently, there is no reliable way to fix or improve this—some models may produce unreadable or blank outputs on first run or after reopening the terminal.

This is a known limitation related to the quantized model formats and their initialization/loading behavior.