Source: harness-hub/Qwen3.8-27B-pi-adapted model card (Apache-2.0 LoRA on Qwen3.8-27B), part of a Hugging Face collection of 11 harness-specific adapters; announced by Hugging Face CEO Clem Delangue on LinkedIn, 2026-11-17. The card says the full training recipe will follow in a release blog post.
What it is
A single LoRA adapter trained with reinforcement learning (async GRPO via TRL) so that Qwen3.8-27B works well as the model behind one specific coding agent, the open-source pi harness. Training data comes from two harness-hub task sets for software-engineering and terminal work. The point is not a better general model but a model fitted to one harness's tools, prompts and API shape.
Results (as reported on the card)
All runs use the pi harness, one attempt per task at temperature 0.8 with thinking off, paired against the base model under pi on the same tasks; intervals are 95% and p-values come from an exact sign test on tasks whose outcome changed.
Terminal-Bench 2.1 (86 tasks): 39.5% to 55.8% (+16.3 points), 16 newly solved and 2 lost, 62% fewer tool calls per task (48% fewer on tasks both solved).
Held-out check: on the 68 Terminal-Bench tasks not used to pick the checkpoint, 39.7% to 55.9%, so checkpoint selection did not drive the gain.
SWE-bench Pro, hardest 51: 29.4% to 62.7% (+33.3 points, interval about plus or minus 14), 18 newly solved and 1 lost, 55% fewer tool calls.
The card is careful in ways launch posts often aren't: paired comparisons, a sign test, a held-out split, and tool calls measured on tasks both models solved, so efficiency gains aren't just an artifact of failing faster.
How it is meant to be run
Load the adapter on the exact base model (it does not transfer to other bases), merge it into the base weights with PEFT for a standalone checkpoint, or serve the base plus LoRA through llama.cpp-style local apps or transformers serve. The card warns that pi must use the chat-completions API: by default pi uses a different API shape, which the adapter was neither trained nor evaluated with.
Why it matters
Harness fit is a large, cheap lever. The same weights plus one adapter roughly doubled hard SWE-bench Pro and cut tool calls in half. Much of what looks like model capability in agent benchmarks is model-harness fit.
Open weights make this possible. Closed models can only be prompted toward a harness; open ones can be trained for it, including by the harness's own users.
Fewer tool calls is a cost and latency win on top of accuracy.
Tradeoffs and failure modes
Overfitting to one harness. An adapter tuned for pi's tools and API shape may get worse in other harnesses or after pi changes its prompts or tool schemas. Version the adapter against the harness version.
API-shape brittleness. A silent config default (the wrong API mode) moves you off the evaluated distribution without an error.
Small benchmark slices. 51 and 86 tasks give wide intervals; the direction is clear, the exact size less so.
Training-set contamination risk if the RL task sets overlap the evaluation suites; the held-out Terminal-Bench split addresses checkpoint selection, not training overlap.
Thinking off, temperature 0.8, one attempt: results may differ in other settings.
Review checklist
Is the adapter pinned to the exact base model and harness version you run?
Is the harness configured with the same API mode and tool schema used in training?
Did you re-run a paired comparison on your own tasks, base versus adapter, in your harness?
Do you track tool calls per solved task, not just pass rate?
What is the plan when the harness changes its prompts or tools: retrain, or fall back to the base?
Related entries
R067 Inkling 2 Small: an open-weight MoE built for agent loops and fine-tuning.
R111 The open-weights gap in November 2026.
R114 AutoRL.
R118 Prefix caching for coding agents.