Latest in AI
Qwen3.8 Pi Adapter Setup: Match the Harness Before Comparing Results
The new pi-specific LoRA has usable local serving paths. Match the base model, API and thinking settings, then evaluate the resulting agent rather than borrowing its benchmark score.
The newly available Qwen3.8-27B-pi-adapted adapter is a rank-32 LoRA trained for the pi coding agent. It modifies Qwen3.8-27B rather than replacing the whole model. The useful question is whether a model adapted to your agent's tools can finish work with less unnecessary effort.
That question needs a specific setup. A checkpoint running in a chat window is not the same system as a coding agent reading files, executing commands and deciding what to do next. I would reproduce the intended connection first, then change one component at a time.
Start with the exact base model
For the llama.cpp route, the official GGUF card provides an approximately 319MB adapter file. That download is not a complete coding model. It must be loaded alongside a GGUF of Qwen3.8-27B.
The card recommends Q8_0 as a starting point and notes that lower-bit quantization moves further from the BF16 weights used in training. It explicitly says the GGUF version has not been re-evaluated through llama.cpp.
Before downloading, budget for the base weights and runtime memory, not just the adapter. Record the base checkpoint, quantization, adapter revision and inference-engine version together. “Qwen with a LoRA” is not enough information for someone else to reproduce your result.
I would avoid combining this trial with a new community model merge. If something breaks, you want a short list of possible causes.
Keep the evaluated serving settings together
The original card's vLLM configuration disables thinking, enables rank-32 LoRA, uses the qwen3_xml tool parser and sets mamba-cache-mode align for prefix caching. Its pi provider configuration selects openai-completions, not the default Responses API.
These settings are a connected recipe, not independent tuning suggestions. Start from the linked configuration and check that the model identifier pi requests matches the identifier your server exposes.
Test the connection in stages. First request plain text. Then ask the agent to read a harmless file and report a fact you can check. Only after those work would I ask it to change a small test project.
Keep server errors alongside the agent transcript. If a request never reaches the model, that is a serving failure, not evidence that the adapter cannot solve the task. I have not run this setup; these are the checks I would use for an initial trial.
Keep local tool access bounded
The pi repository says it has no built-in permission system restricting filesystem, process, network or credential access. It runs with the permissions of the launching process and documents container or sandbox options for stronger boundaries.
Local inference does not change that. A locally hosted model can still issue an unwanted command through a coding harness.
Use a disposable project copy in an appropriately isolated environment, without production credentials or unrelated personal files. Define the task's allowed changes before running it. Preserve the original files so you can inspect the diff rather than relying on the agent's account of what changed.
For a first task, I would choose a small parser bug with an existing failing test. It gives you a concrete acceptance condition and a manageable set of changes to review.
Match the client to the local server
For llama.cpp, Hugging Face's pi-llama extension supplies a provider with dynamic model and context-window discovery. The GGUF card links this route separately from its vLLM setup.
Do not carry a working configuration from one backend into another and assume every field means the same thing. Confirm the endpoint, served model name and usable context before starting a long task.
Once the short test works, repeat it after restarting the server. A setup that only works because of an undocumented session state will be frustrating to maintain.
Separate promising scores from your own evidence
The adapter card reports corrected evaluations after rerunning rollouts affected by tunnel failures. It also says checkpoint selection used SWE-bench Pro and an 18-task Terminal-Bench subset, with a separate remaining-task analysis. Those details matter more than an isolated percentage.
For your comparison, give the base model and adapter the same project, instructions, tools and time budget. Count failed runs and connection failures separately. Compare completed tasks, elapsed time, generated tokens and human corrections.
Then inspect the resulting code. Fewer turns are valuable only when the agent still produces the right change. My preference would be a small collection of repeatable tasks before switching a working coding workflow. The adapter is a promising component, but the thing you deploy is the full agent system.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help AI teams turn model releases and developer tooling into clear, practical technical articles.