# Initial experiences with pi adapter on q4 qwen3.8 27b — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/6950263c9a59fea5

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-19 08:01 ingested 2026-11-20 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1y03xuy/initial_experiences_with_pi_adapter_on_q4_qwen38/)

## retained text



```
Hi! Just wanted to share my short experience with qwen3.8 27b's pi adapter from huggingface.

The use-case I thought it'd fit well was for design with penpot. I'm a dev, but need to do designs sometimes, and I need to make designs for a legacy codebase so it's code + rendered HTML/CSS to penpot design.

I thought this model'd be perfect because penpot MCP takes tons of small tool calls and good code, algebraic and geometric reasoning.

Sooo... how'd it go?

\---

1. Token generation is slower. I don't have metrics, but t/s is clearly slower compared to running the same model without the LORA. I suspect this is due to MTP being less effective.

2. The model understands the pi tools much better. Normally I have to tell it to use codemode and mcp which is frustrating, and then it'd get the first two or three wrong. Not the case here, it hits the ground running.

3. It totally breaks sometimes. Crazy loops where it stops functioning entirely. I suspect this has to do with the fact that I'm applying the lora to a quantized model, my quant is Q4\_K\_S.

4. The thinking off is interesting, because it uses regular text for thinking. Much smaller reasoning traces than I'm used to for the thinking modes. It seems to hold its ground just fine regarding reasoning, but lacking effort modes is not my preferred way of interacting with the model.

Conclusion?

\---

So close. I feel like this is so close to being a straight upgrade to qwen3.8 27b, but the (suspected) MTP slowdown and the lack of ready-to-go quants causing the loop problems make me feel like this is a research preview and not a production ready model.

I will not be replacing my Q4\_K\_S Qwen with this one yet.

Am curious to see if this is where huggingface leaves the models, or if they'll address the problems I've had, it's a very promising prospect to teach a model to learn its harness well.
```



document fetched 2026-11-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1y03xuy/initial_experiences_with_pi_adapter_on_q4_qwen38
