← sources

Qwen3.5 9B running as a fully local AI agent on my phone — including code execution and PDF generation

source: reddit (reddit:LocalLLaMA)published 2026-09-05 15:59ingested 2026-09-06 02:41

retained text

**Running a fully local 9B AI agent on a phone — LLM, tools, code execution, and file generation**  
I’ve been experimenting with how far a **fully local LLM agent** can go on a phone.  
This demo runs **Qwen3.5 9B IQ4\_XS locally on-device** using TensorSharp.  
I give it a very simple prompt:  
“Please get 10 stocks with most gains today.”  
From that, the local agent:  
reasons about how to solve the task  
generates Python code  
executes the code locally  
retrieves the requested public stock-market data  
returns the result  
Then I ask:  
“Convert it to a PDF.”  
There is no hard-coded PDF workflow.  
The agent discovers the appropriate **Skill**, reads its SKILL.md, decides how to perform the task, writes the necessary code, executes it, generates the PDF, and returns the file — all on the phone.  
So what is running locally is more than just token generation:  
**Local LLM → reasoning → Skills → tool use → code generation → local execution → file generation**  
The model is **Qwen3.5 9B IQ4\_XS**, and inference is handled locally by **TensorSharp**, my open-source GGUF inference/runtime project.  
The part I find most interesting is privacy.  
Your **prompts, conversation history, local files, and private on-device data do not need to be sent to a cloud LLM provider**.  
The phone only accesses the network when the actual task requires external public information — in this example, stock-market data.  
So the AI itself can stay local while still acting as a useful agent.  
I think this becomes especially interesting on mobile devices, where local inference can provide:  
privacy  
offline capability  
no per-token API cost  
lower dependence on cloud services  
access to local files and device-side workflows without uploading them  
TensorSharp repo:  
https://github.com/zhongkaifu/TensorSharp  
I’m curious how people here see the future of **on-device local agents**.  
Do you think 7B–12B quantized models are already good enough for useful autonomous workflows on phones, or are we still mostly waiting for better models and more mobile memory bandwidth?

document fetched 2026-09-06 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1w84k0z/qwen35_9b_running_as_a_fully_local_ai_agent_on_my