Used PI agent + Qwen 3.8 27b ( q4xl with mtp ) to port Mira_ tts - WORKING!

[mirek190]

I am shocked the small local model like Qwen 27b 3.8 was able to DO THAT.

Used PI agent + qwen 3.8 27b with cache Q8 with the MPT for increase speed , context 125k.

Qwen used over 3m tokens and was working over 4 hours ( MPT allowed to do that 2x faster )
BUT IS FULLY PORTED AND WORKING.

Generated audio
https://huggingface.co/mirek190/audio.cpp/blob/main/out/out.wav

reference
https://huggingface.co/mirek190/audio.cpp/blob/main/out/demo_1_man.wav

[IIIIIllllIIIIIlllll]

yes, Qwen3.8 27B is real powerful

[mirek190]

It is insane .... I did not expect 27b model could do that.

[IIIIIllllIIIIIlllll]

Qwen3.8 27B performs quite well, close to Opus 4.6. The biggest drawback is that the chain‑of‑thought is overly verbose — after using it for a while and checking the token statistics, I found that the output tokens exceeded half of the input tokens. On consumer‑grade hardware, the speed isn't particularly fast either, making it too slow for complex model porting tasks.

[mirek190]

IIIIIllllIIIIIlllll

What I learn working with PI agent with qwen 3.8 27b within 2 days.
If you want increase speed 2x - 3x use MTP module. That is really giving boost but is reducing context ( for me from 180k to 120k with RTX 3090 )

Yes on xhigh thinking setting is thinking like a crazy that is true ... but only in the first phase ( making plan and analyze that plan ).
After that the first phase is thinking not too much actually.

For instance I asked later to add implementation for gguf Q8 and bf16.
The agent added that implementation within 5 mintues ( preparing new guufs and change code )

[0xShug0]

@mirek190 Amazing! Looks like the local setup is sufficient for porting small, edge-friendly models suggested by users. I’m curious how good it is at optimization. I’ve never really managed to get Codex to automatically find meaningful optimization opportunities It tends to get stuck on hair-splitting details.

[mirek190]

@mirek190 Amazing! Looks like the local setup is sufficient for porting small, edge-friendly models suggested by users. I’m curious how good it is at optimization. I’ve never really managed to get Codex to automatically find meaningful optimization opportunities It tends to get stuck on hair-splitting details.

I think because you are using VS code.
Try codex-cli for it as it has built-in many good tools / skills and just working much better with own GPT agents.

Still testing qwen 3.8 27b via PI agent and was able improve performance for CPU backend from 235 seconds to 28 seconds already ....

[0xShug0]

Cool will check it out later!

[mirek190]

Cool will check it out later!

Qwen 3.8 27b -> yes is possible to improve performance even far more with that model ( testing for cpu ) - used around another 1m tokens to increase performance almost x6 times on cpu

from 235s -> 125s -> 28s-> now to 5.3s and still trying to improve kernel even more ....

that is doing a small local model .... incredible

[mirek190]

...and got again 2x faster on cpu - that cost me only 500k more tokens / some 50 minutes of work

Qwen 3.8 27b with PI agent is certainly capable to optimize models.