A 638-megabyte adapter released on Hugging Face lets a 27-billion-parameter model match its own high-effort reasoning scores on coding benchmarks while generating roughly a quarter of the tokens. The adapter, published by the harness-hub team on November 16, is built for a single coding agent called Pi.
What the Adapter Is and Who It Is For
The file is a LoRA — a low-rank adaptation — that sits on top of Qwen/Qwen3.8-27B, an open-weight model released by Alibaba’s Qwen team. Instead of fine-tuning the whole model for general chat, the harness-hub researchers trained this adapter to work inside one specific harness: the Pi coding agent, an MIT-licensed command-line toolkit. Developers who already run Qwen3.8-27B on their own hardware can swap in the adapter without changing the base model weights. The adapter is 638 megabytes; a merged BF16 checkpoint comes to about 54 gigabytes. It runs on vLLM with LoRA enabled or on llama.cpp with a GGUF adapter on a Q8_0 base. The license is Apache-2.0.
How It Differs From General Fine-Tunes
Most published adapters target broad instruction-following or chat. This one targets a single agent’s tool schema, compaction rules and permission flows. The harness-hub team argues that the harness should stay fixed and the model should adapt — one small LoRA per agent, hot-swapped at runtime. A sibling adapter for the OpenCode agent, released the same week, reports 51.0 percent on the same SWE-bench Pro hard-51 subset, up from 35.3 percent for the base model in that harness. The Pi adapter is not meant to improve Qwen3.8-27B on other agents or on general benchmarks.

Benchmark Gains Inside Pi
On the 51 hardest tasks of SWE-bench Pro, the base model with thinking disabled solves 39.2 percent at pass@1. The Pi adapter raises that to 64.7 percent, a gain of 25.5 percentage points. It solves 15 tasks the base model could not and loses two it previously solved. On Terminal-Bench 2.1, pass@1 rises from 39.5 percent to 55.8 percent, a gain of 16.3 points, with 16 newly solved tasks and two lost.
The adapter matches the base model running at its highest reasoning effort (xhigh) on Terminal-Bench 2.1 — 55.8 percent versus 56.6 percent — and recovers 25 of the 35-point gain that thinking provides on SWE-bench Pro (64.7 percent versus 74.5 percent). All evaluations used one attempt per task, a 100-step budget, thinking off, temperature 0.8, top-p 1.0, text-only serving with the vision tower disabled, and identical settings for base and adapter.
Efficiency Gains: Fewer Tokens, Fewer Turns
The larger story is how much less work the adapter does. Across all evaluated tasks, it makes 55 percent fewer tool calls on SWE-bench and 62 percent fewer on Terminal-Bench. On tasks both models solve, the adapter needs a median of 23 agent turns on SWE-bench Pro against 59 for the base model, and 8 turns against 18 on Terminal-Bench 2.1. The adapter writes more visible reasoning between tool calls, about 160 tokens per turn versus roughly 10 for the base model, but it spends far fewer turns overall.
On one SWE-bench Pro task — fixing a human_to_bytes function in Ansible — the base model searches and re-runs snippets until it hits the 100-step limit. The adapter finds the function, edits it, runs the tests and fixes the last case in eight tool calls.
Training Method: AsyncGRPO Inside the Unmodified Harness
The adapter was trained with AsyncGRPO, a reinforcement-learning algorithm from the TRL library, using eight rollouts per task at temperature 0.8. Training tasks mixed 750 scale800 tasks with 300 TermGrade tasks. The reward combined verifier correctness with a tool-efficiency bonus weighted at 0.1. Each rollout had a budget of up to 50 agent steps and 1,800 seconds, with thinking disabled. Every tool call passed through safeagents, a rule-based proxy that blocks calls such as fetching benchmark answers from the Hub; blocked rollouts score zero, pushing the model away from reward hacking. Training ran on two NVIDIA B200 GPUs and stopped at step 191 of a planned 1,000. The adapter applies to the query, key, value, output, gate, up and down projections across all 64 layers of Qwen3.8-27B, using LoRA rank 32 with alpha 64.
Checkpoint Selection and the Held-Out Evidence
The checkpoint at step 160 was chosen by watching scores on SWE-bench Pro and on an 18-task slice of Terminal-Bench 2.1 during training. That means SWE-bench Pro was not a held-out test. On the 68 Terminal-Bench tasks outside that selection slice, performance still rose from 39.7 percent to 55.9 percent — the cleanest evidence in the model card that the improvement carries beyond the tasks used for selection. With only 51 tasks in the SWE-bench Pro hard subset, the model card notes that the true gain could plausibly sit anywhere from about 11 to 40 points.
Deployment and Caveats
To serve the adapter with vLLM, launch Qwen/Qwen3.8-27B with --enable-lora --max-lora-rank 32 and point Pi at the local endpoint. For llama.cpp, load the GGUF adapter on a Q8_0 base GGUF. On macOS, Llama.app can register the adapter with a 32,768-token context. Merging the adapter with PEFT on CPU produces the 54-gigabyte BF16 checkpoint that can be served without LoRA flags.
The results are developer-reported on the model card, not independently replicated. The card explicitly notes small evaluation subsets, Pi-only applicability, and that the checkpoint was selected using tasks that also served as evaluation tasks. An earlier social-media post quoted a baseline of 29.4 percent on SWE-bench Pro; the model card now reads 39.2 percent, and the baseline correction has been noted by third-party coverage. The numbers are not comparable with the Qwen3.8-27B model card, which uses different harnesses (Terminus, Claude Code), maximum-effort thinking, larger budgets and the full SWE-bench Pro suite.
Sources
- harness-hub/Qwen3.8-27B-pi-adapted · Hugging Face
- Qwen3.8-27B’s pi adapter reports better coding scores—and fewer trips to the toolbox
- New LoRA Adapter Lifts Qwen3.8 Coding Success From 39% to 65%
- OpenCode LoRA for Qwen3.8-27B: Benchmarks and Setup | explainx.ai Blog
- Qwen3.8 27B Pi Adapter: Benchmarks, Training and Local Setup | AgentsPulse
Photo: Bernd 📷 Dittrich / Unsplash (Unsplash License)





