Results

Table with columns: Benchmark, Qwen3.8-27B, Qwen3.8-27B, xhigh thinking, Qwen3.8-27B-pi-adapted, Qwen3.8-27B-pi-adapted, xhigh thinking| Benchmark | Qwen3.8-27B | Qwen3.8-27B, xhigh thinking | Qwen3.8-27B-pi-adapted | Qwen3.8-27B-pi-adapted, xhigh thinking |
|---|
| SWE-bench Pro (hard 51) | 39.2% | 74.5% | 64.7% | 74.5% |
| Terminal-Bench 2.1 | 41.0% | 54.2% | 56.6% | 56.6% |
On Terminal-Bench 2.1 the four runs are compared on the 83 tasks all of them graded, so the base model and adapter scores differ slightly from the paired results below, which use the 86 tasks those two graded. "Thinking" is the chat template's
thinking mode at its default reasoning effort (xhigh).
One attempt per task, thinking off, paired against the base model under pi on the same tasks. The interval is 95%; the
p-value is an exact sign test on the tasks that changed outcome.
Table with columns: Benchmark, Tasks, Qwen3.8-27B, Qwen3.8-27B-pi-adapted (step 160), Difference (points), Newly solved / lost| Benchmark | Tasks | Qwen3.8-27B | Qwen3.8-27B-pi-adapted (step 160) | Difference (points) | Newly solved / lost |
|---|
| SWE-bench Pro (hard 51) | 51 | 39.2% | 64.7% | +25.5 ± 14.2 | 15 / 2 (p = 0.0023) |
| Terminal-Bench 2.1 | 86 | 39.5% | 55.8% | +16.3 ± 9.1 | 16 / 2 (p = 0.0013) |
| Terminal-Bench 2.1, the 68 tasks not used to pick the checkpoint | 68 | 39.7% | 55.9% | +16.2 ± 10.6 | 13 / 2 (p = 0.0074) |
- SWE-bench Pro (hard 51): the 51 hardest tasks of SWE-bench Pro, 2 h and 100 agent steps per task.
- Terminal-Bench 2.1: all 89 tasks of Terminal-Bench 2.1, 1 h and 100 agent steps per task.
Step 160 was picked by looking at the checkpoints' scores on SWE-bench Pro and on an 18-task subset of Terminal-Bench
2.1, tracked during training.
Medians per task, this adapter / the base model, over the tasks both solved: same tasks, same outcome, so the numbers
measure only how much work each needed. Generated tokens are everything the model writes (visible content, reasoning and
tool calls); content is the visible text between tool calls.
Table with columns: Tasks, Turns, Content tokens, Reasoning tokens, Generated tokens | Tasks | Turns | Content tokens | Reasoning tokens | Generated tokens |
|---|
| vs the base model | | | | | |
| SWE-bench Pro | 18 | 23 / 59 | 3.5k / 0.6k | 0 / 0 | 6.9k / 9.6k |
| Terminal-Bench 2.1 | 32 | 8 / 18 | n/a | 0 / 0 | 1.5k / 2.9k |
| vs the base model with thinking | | | | | |
| SWE-bench Pro | 27 | 27 / 74 | 3.5k / 1.3k | 0 / 34.8k | 8.8k / 46.7k |
| Terminal-Bench 2.1 | 35 | 8 / 18 | 0.9k / 0.8k | 0 / 9.9k | 1.4k / 11.6k |
- Against the base model, the adapter writes short reasoning between tool calls (about 160 tokens per turn against
about 10) and needs less than half the turns.
- Against the base model with thinking, nearly all the difference is reasoning: the adapter reaches the same answers
without it.
- Content for the base model on Terminal-Bench 2.1 is not available: its run's traces were not kept.
Table with columns: SWE-bench Pro (hard 51), Terminal-Bench 2.1 | SWE-bench Pro (hard 51) | Terminal-Bench 2.1 |
|---|
| This adapter vs the base model with thinking (tasks only one solves) | 6 / 11 (p = 0.33) | 12 / 10 |
| Generated tokens per task (median), base model with thinking / this adapter | 48.8k / 9.4k* | 16.5k / 3.7k |
| Model calls per task (median), base model with thinking / this adapter | 61 / 21 | 19 / 9 |
| This adapter with thinking vs without (tasks only one solves) | 8 / 3 (p = 0.23) | 9 / 9 |
On Terminal-Bench 2.1 the adapter without thinking is level with the base model thinking. On SWE-bench Pro, thinking
takes the base model further (74.5% against the adapter's 64.7%; not significant on 51 tasks), at about 4.5× the
tokens and 3× the model calls. With thinking on, the adapter reaches the same 74.5%. Thinking runs use the template's
default reasoning effort (xhigh) and up to 32k output tokens per model call.
* Counted from the trajectories: the server's token counts are missing for a third of the SWE-bench Pro thinking
rollouts, mostly long ones.
- Correction, thinking on. An earlier version of this card reported 64.7% (base model) and 66.7% (this adapter) on
SWE-bench Pro with thinking on. pi then capped each model call at 16k output tokens, which cut long reasoning short.
We reran both with the 32k cap used everywhere else; the numbers above (74.5% for both) are the corrected ones.
Thinking-off results were not affected.
- Correction, tunnel. An earlier version of this card reported 29.4% → 62.7% on SWE-bench Pro. Nine of the base
model's rollouts there had failed because the public tunnel between the sandboxes and the model server went down
mid-run. We now check every rollout for this and reran all affected ones, on both sides; the numbers above are the
corrected ones.
- Budget and time. The base model often works until it runs out of its 100 model calls: on SWE-bench Pro, 20 of its
31 failed attempts ended this way, against none for this adapter. With thinking on, Terminal-Bench sessions often hit
the one-hour limit (18 of the base model's 39 failed attempts). Limits are the same for every run, so the comparisons
are fair, but part of the adapter's gain is finishing within them. pi never ran out of the 131,072-token context.
- Network access. The sandboxes had internet access. Calls outside the task were almost all dependency lookups (a
module path or version a task needed); we found no attempt by this adapter to look up a task's upstream fix.
- Test files. Agents, the base model and this adapter alike, sometimes write or edit test files while working,
mostly their own scratch scripts to check a change. This cannot affect a grade: SWE-bench Pro's grader resets all test
files and applies the official tests before running them, and Terminal-Bench's tests are copied in only at grading.
We found no tampering with tests or the verifier.
- Hidden reasoning. The adapter's visible reasoning is plain text:
<think> tags appear in its output in fewer than
1 turn in 1,000, and with thinking off the template never opens a thinking block.
Evaluation trajectories
Every evaluation attempt behind this card, solved or not, is in the
harness-hub/harness-adapters-trajectories
bucket: base model and adapter, thinking on and off, on both benchmarks, with each attempt's grade, trajectory, harness
log and verifier output, scrubbed of secrets. The files carry canary strings: keep them out of training data.
Use with pi
1. Serve the adapter with vLLM
vllm serve Qwen/Qwen3.8-27B \ --enable-lora --max-lora-rank 32 \ --lora-modules qwen3.8-pi=harness-hub/Qwen3.8-27B-pi-adapted \ --language-model-only --enable-prefix-caching --mamba-cache-mode align \ --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 \ --default-chat-template-kwargs '{"enable_thinking": false}' \ --max-model-len 131072
This is the configuration of the evaluations. Thinking off is what the adapter was trained for: with it on, it
generates about 4× more tokens for at most a few more solved tasks. Prefix caching on this hybrid model needs
--mamba-cache-mode align; without it, every agent turn re-reads its whole history.
2. Point pi at it
Install pi (Node.js 22.19 or newer):
npm install -g --ignore-scripts @earendil-works/pi-coding-agent
Add the server to ~/.pi/agent/models.json:
{ "providers": { "local": { "baseUrl": "http://localhost:8000/v1", "api": "openai-completions", "apiKey": "dummy", "models": [{ "id": "qwen3.8-pi" }] } }}
"api": "openai-completions" matters: without it pi uses the Responses API, which this adapter was not trained or
evaluated with.
pi --provider local --model qwen3.8-pi --print "Find out why the tests in tests/test_parser.py fail and fix the bug."
transformers serve is a lightweight OpenAI-compatible
server, for local use and experimentation (vLLM is faster under load). It handles Qwen's tool calls out of the box.
Serve the merged checkpoint from Merge the adapter into the base model:
pip install "transformers[serving]"transformers serve ./Qwen3.8-27B-pi-merged --dtype bfloat16 --continuous-batching --reasoning off
--reasoning off keeps thinking off, as in training. The server listens on http://localhost:8000, so the pi
configuration from the vLLM section works as is, with the model id set to the name the server
serves, here ./Qwen3.8-27B-pi-merged:
pi --provider local --model ./Qwen3.8-27B-pi-merged --print "Find out why the tests in tests/test_parser.py fail and fix the bug."
Use with pi (llama.cpp)
The GGUF version of Qwen3.8-27B-pi-adapted is in harness-hub/Qwen3.8-27B-pi-adapted-GGUF. It runs on top of a GGUF of the base model:
hf download ggml-org/Qwen3.8-27B-GGUF Qwen3.8-27B-Q8_0.gguf --local-dir .hf download harness-hub/Qwen3.8-27B-pi-adapted-GGUF Qwen3.8-27B-pi-lora-F16.gguf --local-dir . 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
Then point pi at it with the pi-llama extension:
pi install git:github.com/huggingface/pi-llamapi --model llama-cpp/qwen3.8-pi
The adapter was trained on the BF16 weights: the closer the base GGUF is to them, the closer you get to the trained
behavior. Q8_0 is a good default; lower-bit quants drift further. The GGUF version has not been re-evaluated.
Run it with Llama.app (macOS)
-
Install Llama.app and launch it once. On first launch it creates ~/.config/llama/models.user.ini.
-
Download the adapter and register the model:
LORA=$(hf download harness-hub/Qwen3.8-27B-pi-adapted-GGUF Qwen3.8-27B-pi-lora-F16.gguf)cat >> ~/.config/llama/models.user.ini <<EOF [harness-hub/Qwen3.8-27B-pi]hf-repo = unsloth/Qwen3.8-27B-GGUF:Q4_K_XL ; This is just an example but you can also use any GGUF that is derived from Qwen3.8-27B lora = $LORAtemp = 0.7ctx-size = 32768EOF
hf-repo sets the base model: the app downloads it on first use, about 17 GB. To use another quant of the same base, change the part after the colon, for example :Q8_0.
-
Restart Llama.app and choose harness-hub/Qwen3.8-27B-pi from the model list.
You can also use it through the app's OpenAI-compatible server. Llama.app shows the port; the example below uses 9931:
curl localhost:9931/v1/chat/completions -d '{"model":"harness-hub/Qwen3.8-27B-pi","messages":[{"role":"user","content":"hi"}]}'
The adapter must be loaded on top of Qwen3.8-27B. It does not work with other base models.
Merge the adapter into the base model
The adapter can be merged into the base weights with PEFT. The result is a regular Qwen3.8-27B checkpoint (about 54 GB
in BF16) that any engine serves without LoRA support. Merge on CPU if the GPU is needed for serving afterwards:
import torchfrom peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoTokenizer base = AutoModelForImageTextToText.from_pretrained("Qwen/Qwen3.8-27B", dtype=torch.bfloat16, device_map="cpu")model = PeftModel.from_pretrained(base, "harness-hub/Qwen3.8-27B-pi-adapted").merge_and_unload()model.save_pretrained("Qwen3.8-27B-pi-merged")AutoTokenizer.from_pretrained("harness-hub/Qwen3.8-27B-pi-adapted").save_pretrained("Qwen3.8-27B-pi-merged")
Then serve it like the base model, without the LoRA flags:
vllm serve Qwen3.8-27B-pi-merged --served-model-name qwen3.8-pi \ --language-model-only --enable-prefix-caching --mamba-cache-mode align \ --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 \ --default-chat-template-kwargs '{"enable_thinking": false}' --max-model-len 131072
The merged checkpoint can also be converted to GGUF and served with llama.cpp the same way.
Training
Table | |
|---|
| Method | AsyncGRPO (TRL, experimental), LoRA |
| Environment | Harbor tasks served by OpenEnv, rollouts in HF sandboxes, tool calls screened by safeagents |
| Harness | pi |
| Tasks | 750 scale800 tasks interleaved with 300 TermGrade tasks; 1,919 task groups trained up to this checkpoint |
| Reward | the task verifier's correctness (falls back to reward), plus a tool-efficiency bonus of weight 0.1 |
| Steps | 160 (this checkpoint); the run was stopped at step 191 of 1000 planned |
| Learning rate | 2e-5 |
| LoRA | r = 32, alpha = 64, on q, k, v, o, gate, up and down projections |
| Rollouts | 8 per task, temperature 0.8, thinking off, up to 50 agent steps and 1,800 s per rollout |
| Batch | 16 micro-batches of up to 65,536 tokens per step, rollouts up to 16 policy versions stale |
| Hardware | 2× NVIDIA B200 per harness: vLLM on one, the trainer on the other |
The training script is
examples/async_grpo_harbor in TRL.
Below is an animated training curve of the evolution of the number of tool calls on the train set:

For a full write-up, please refer to the release blogpost.