Qwen3.8-27B-pi

Qwen3.8-27B-pi

A 27B coding model tuned for the Pi agent framework, trained to read a repository, edit files, run tools and use their feedback, with adjustable reasoning effort. For developers who run coding agents.

Made by bytkim on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8-27B-pi is a fine-tune of Qwen3.8-27B for coding work inside the Pi agent framework. Training centered on a repeating cycle: look through a codebase, change files, run tools, then respond to what they report. The goal is more finished work with less generated text. You can set the reasoning effort, from low to xhigh.

Who it suits: Developers who run coding agents on their own hardware and want a model shaped for tool-driven coding sessions. It is aimed at the Pi framework, so it fits best when you use that.

Industry | Who uses it | How they use it | Example
Software | Developers | Let an agent edit code and check its own results | Ask it to fix a failing test and rerun the suite
AI agents | Agent builders | Run a coding model inside the Pi framework | Use medium effort for routine edits, xhigh for hard bugs
DevOps | Engineers | Handle terminal-style tasks that need tool feedback | Work through a task with repeated command runs
Research | ML practitioners | Compare reasoning effort against output length | Test low, medium and xhigh on the same task

You can skip it if you do not use the Pi framework or need a general chat model. The card describes it as built for coding work.

Where to use it:

• The Pi agent framework, which the model is tuned for
• Python with transformers on Windows, macOS and Linux, for the BF16 weights
• FP8 and GGUF copies, which the card links for other hardware

Use case | Who | What they get back
Repository edits | Developers | Code changes made through tool calls
Debugging | Engineers | Fixes guided by tool feedback
Effort tuning | Agent builders | A choice between quick answers and deeper reasoning

How to set it up: You need Python and enough memory for a 27B model, about 55.6 GB in BF16 or about 13.9 GB at 4-bit by estimate.

• Python on Windows, macOS or Linux
• The transformers library, or a GGUF-capable tool for the quantized copy
• The Pi agent framework, for the coding workflow

1. Install Python if you do not have it.
2. Install the Hugging Face hub tool.
3. Download the model files.
4. Load them in your serving tool or in the Pi framework.

pip install -U huggingface_hub
hf download bytkim/Qwen3.8-27B-pi

How to use it: The quickstart section is cut off in the part shown here, so take the exact run steps from the full card. In use, you give the model a coding task in the Pi framework and pick a reasoning effort level.

Example: You ask it to find why a test fails in a repository. It reads the code, makes a change, runs the tests, and uses the tool output to decide its next step. Medium effort handles routine fixes, and xhigh suits harder debugging.

For newcomers and experts: A framework is the program that lets a model read files and run commands. This model has practiced that loop. A first small step is to download the files and read the card's quickstart before loading it.

Its first training stage used supervised learning on Pi sessions that had been filtered to keep the successful ones. A second stage used reinforcement learning with GRPO and a reward for efficient reasoning, aimed at low and medium effort. The card compares it with the base model on Terminal-Bench 2.1, GPQA Diamond and SciCode. These are the authors' selected results.

Read the repository
Make a change
Run tools and check the result
Use the feedback for the next step

Title: Fewer output tokens than the base model, in the card's selected results
Terminal-Bench, Pi medium against Base xhigh: 41%
SciCode at xhigh, same or better score: 23%

Good to know:

• On GPQA Diamond the card says the base model keeps the edge at medium effort, while Pi has the best xhigh score.
• On SciCode, Pi at medium effort scores higher but uses more tokens.
• The weights are 27.8B parameters, with FP8 and GGUF copies listed beside the BF16 one.
• License: apache-2.0.
• Cost to run: none named, you use your own hardware.
• Data it reaches: files downloaded from Hugging Face, plus whatever repository and tools you give the agent.
• Not for: the card names no limits beyond its focus on coding in the Pi framework.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,857
• Likes: 87
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 13
• Fine-tunes: 2
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, multi-token-prediction, speculative-decoding, vision, image, multimodal, qwen, qwen3.8, qwen38

Links

• Model: https://huggingface.co/bytkim/Qwen3.8-27B-pi
• Files: https://huggingface.co/bytkim/Qwen3.8-27B-pi/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download bytkim/Qwen3.8-27B-pi

Then follow the usage in its model card. Nothing in this kit runs by itself.