Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
Back to feed
bytkim 
posted an update Jul 9
Post
318
Qwen3.6-27B-pi-tune v2 is coming plus its sibling 35B-A3B variant

I just wanted to share an update on the progress of future releases for Qwen3.6-pi-tuned family models.

Both 27B and 35B models are now unified under native think/no-think functionality.

The biggest lesson I started to learn after reviewing many suggestions: Fine-tuning for local open-weight agents isn't just about raw coding capability or benchmark numbers. Harness fluency, tool-calling, validation loops, and user-facing behavior matter just as much, sometimes more.

That insight and philosophy is exactly what v2 is based on.

Although it hasn't even been a month since the original release I wanted to get out the 35B-A3B variant as soon as possible due to popular demand.

With the current and upcoming releases of a new class of Agentic LLM's (Fable, GPT5.6, etc) expect v3 to be the best yet.

For everyone already running the original: what's it doing well, and what makes you reach for a different model? These suggestions help shape future releases.

https://huggingface.co/collections/bytkim/qwen36-pi-tune

PS: Heres a sneak peek at the newest version with the new 35B-A3B variant performing a one shot task. 😁
akierum

Your model is best in class.

leaderboard

Waiting for Qwen 3.8 27B pi thinking model too.
Your models are the only ones passing that codeneedle test no making empty space errors on first row.

I would suggest releasing the XL versions too as BF16 is out of reach for everyone, and keep up the good work.

·
bytkim
•
edited 27 minutes ago

Hi akierum,

Thank you, I appreciate the kind words.

I stepped back from pi-tune for a while, but I am happy to confirm an upcoming Qwen3.8-27B-pi-tune. I aim to complete the full release by next Friday, 09/11/26.

I am trying my best to adhere to Qwen3.8's canonical reasoning template and intended reasoning behaviors with the target of harness/tool fluency and token efficiency. The upcoming release will be one model with FP8, GGUF, NVFP4, etc.... I will be including a diverse set of benchmarks and (hopefully) an interesting accompanying write-up.

As for the planned Qwen3.6-35B-A3B, I spent a lot of time training it and wasn't happy with it. I decided not to ship it as I couldn't recommend it over the 27B, and it exhibited functional and behavioral concerns. I've since concluded that future MoE model releases from the Qwen family will require a more planned and informed approach as to not degrade the base model.

I'm finalizing 3.8-27B now and have interesting plans for 3.5-9B afterwards.

As for the XL quants, I'll make sure they're in this release. And thanks again for the evaluation data. If anything else turns up, feel free to send it over, I'm always glad to see data from real use.
In this post