Qwen3.8-27B pi adapter: what the coding-agent scores show
A new LoRA tunes Qwen3.8-27B for the Pi coding agent. What its scores show and miss, the files you need, and the two settings that matter.
· Updated · 3 min read
A new LoRA adapter tunes Qwen3.8-27B to work as the model behind the Pi coding agent, and its makers report a 16-point gain on Terminal-Bench 2.1 with thinking turned off, on the adapter's model card (read 18 November 2026). The Qwen3.8-27B pi adapter was published on Hugging Face by the harness-hub organisation on 16 November, under Apache 2.0, with a GGUF version for llama.cpp. Every score so far is the publisher's own, and it targets thinking-off use, not a higher ceiling.
What is the Qwen3.8-27B pi adapter?
It is a rank-32 LoRA trained with reinforcement learning (TRL's AsyncGRPO) on software-engineering and terminal tasks, with Pi as the harness, according to the card. The card says training stopped at step 191 of a planned 1,000 and the published checkpoint is step 160. It only works on top of Qwen/Qwen3.8-27B, a dense, Apache-licensed vision-language model.
How much better is it, and how far do the numbers go?
Under Pi, one attempt per task at temperature 0.8, thinking off, the corrected model card (read 18 November 2026) reports:
| Benchmark (publisher's run) | Base Qwen3.8-27B | With adapter | Median turns, tasks both solved |
|---|---|---|---|
| Terminal-Bench 2.1, 86 tasks | 39.5% | 55.8% | 18 → 8 |
| Terminal-Bench 2.1, 68 tasks not used to pick the checkpoint | 39.7% | 55.9% | not given |
| SWE-bench Pro, hardest 51 | 39.2% | 64.7% | 59 → 23 |
The card says nine base-model runs had failed when the link to the model server dropped; affected runs were repeated on both sides.
The same card (18 November 2026) also compares the adapter with the base model thinking at its default xhigh effort: 64.7% for both on SWE-bench Pro, and 56.6% against 54.2% on Terminal-Bench 2.1 (on the 83 tasks all runs graded), with about 4.5 times fewer generated tokens for the adapter. Turning thinking on for the adapter added nothing in its tests.
Three cautions. The SWE-bench Pro set was also used to choose the checkpoint, so only the 68-task row is held out. The 86-task row is the tasks both runs graded, out of the dataset's 89. And no independent run had been published as of 18 November that we could find.
For context, Qwen's own card (read 18 November 2026) lists 73.0 for the base model on Terminal-Bench 2.1 with the Terminus harness; it does not state the thinking setting, and thinking is on by default. The adapter's card says the two are not comparable: Qwen's use other harnesses, maximum-effort thinking and larger budgets. Our reading is that the adapter's case is thinking-level accuracy under Pi at a fraction of the tokens, not a higher ceiling.
What it means if you run models yourself
The GGUF route needs a Qwen3.8-27B base file plus the adapter. In the ggml-org GGUF folder (file list read 18 November 2026), Q8_0 is 28.6 GB and Q4_K_M 19.4 GB; the adapter GGUF is 0.32 GB. Add memory for the context you choose. The card's llama.cpp command is below; it recommends Q8_0, because the adapter was trained on BF16 weights, and says the GGUF version has not been re-evaluated.
llama serve -m Qwen3.8-27B-Q8_0.gguf --lora Qwen3.8-27B-pi-lora-F16.gguf \
--chat-template-kwargs '{"enable_thinking": false}' -c 65536 --port 9931 --alias qwen3.8-pi
New to the runtime? Our llama.cpp 1.0 brief covers what changed for llama-server users, and the best local LLM per memory tier helps judge whether a 27B fits your machine.
Using it from your other machines
Corewire starts Pi with thinking off and over chat completions, the two settings this adapter expects. Serve the adapter on the desktop with the graphics card, adopt the server with Add endpoint in the window's Runtimes panel, and drive Pi there from your laptop on the train: the sessions guide shows how.