Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Qwen3.8-27B-pi-adapted

Hugging Face's RL-trained LoRA turns Qwen3.8-27B into a much better pi coding agent without thinking mode, at a quarter of the tokens.

Reviewed by ReviewRamble editors · Updated

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.9/5

A small, well-documented adapter that makes Qwen3.8-27B a far better pi agent with thinking off, and a useful recipe even if you never run pi.

What is Qwen3.8-27B-pi-adapted?

Qwen3.8-27B-pi-adapted is a LoRA adapter for Alibaba's Qwen3.8-27B, trained with reinforcement learning to act as the model behind pi, an open-source terminal coding agent. It was published in November 2026 by harness-hub, a Hugging Face organisation whose launch post is written by Hugging Face engineers, as one of three per-harness adapters (pi, Hermes Agent and OpenCode) plus a merged adapter that runs under all three.

The adapter does not change the base weights. It adds rank-32 LoRA layers on the attention and feed-forward projections, trained with TRL's experimental AsyncGRPO trainer inside the real pi harness on 750 scale800 software tasks and 300 TermGrade terminal tasks, on two NVIDIA B200 GPUs. You load it on top of Qwen3.8-27B in vLLM or llama.cpp, or merge it into a standalone checkpoint of about 54GB in BF16. Every evaluation trajectory is published.

Is Qwen3.8-27B-pi-adapted worth it?

This adapter shows how much a model gains from being trained inside the agent harness it will actually run in. Under pi with thinking off, it lifts Qwen3.8-27B from 39.2% to 64.7% on the 51 hardest SWE-bench Pro tasks, with 15 tasks newly solved and 2 lost, and from 39.5% to 55.8% on 86 Terminal-Bench 2.1 tasks. Both changes pass a sign test (p = 0.0023 and 0.0013), and the Terminal-Bench gain holds at 16.2 points on the 68 tasks not used to pick the checkpoint.

The bigger finding is efficiency. With thinking off, the adapter matches the base model thinking at its highest effort on Terminal-Bench 2.1 (56.6% against 54.2%) and gets 25 of thinking's 35 points on SWE-bench Pro (64.7% against 74.5%), while generating about 4.5 times fewer tokens per task. On tasks both solve, it needs 23 turns where the base model needs 59 on SWE-bench Pro, and 8 against 18 on Terminal-Bench.

The card is unusually candid. It records two corrections to its own numbers, one caused by a 16k output cap that cut off long reasoning and one by a network tunnel that failed during nine base-model runs, and it notes that part of the gain comes from finishing inside the 100-step budget that the base model often exhausts. All trajectories, graded or not, sit in a public bucket.

The catch: it is narrow. The gains are measured only under pi, with text only, temperature 0.8 and the OpenAI completions API; the card warns pi's default Responses API was never trained or tested. The GGUF version has not been re-evaluated, lower-bit base quantisations drift further from the trained behaviour, and training stopped at step 191 of 1,000 planned. With thinking on, it reaches the same 74.5% as the base model, so it adds nothing for people who already pay for long reasoning.

Our take: if you run pi on your own GPU, this is a free upgrade that turns a slow thinker into a quick finisher; if you use another harness, the sibling adapters and the published TRL recipe are the more interesting part of the release.

Where it earns its place

  • A developer who runs pi against a local model on a workstation GPU

    Fix failing tests and small bugs without paying for API tokens

    Serve Qwen3.8-27B with vLLM and `--lora-modules qwen3.8-pi=harness-hub/Qwen3.8-27B-pi-adapted`, thinking off, and point pi at it with the openai-completions API. Sessions finish in about half the turns of the base model.

  • A Mac developer using Llama.app

    Run a stronger local agent from a laptop

    Register a Qwen3.8-27B GGUF (Q8_0 is the card's default; Q4_K_XL is about 17GB) with the F16 LoRA file in models.user.ini, then use the pi-llama extension. The GGUF path has not been re-benchmarked.

  • An ML team at a company with its own internal agent harness

    Train a cheap adapter that fits their harness and task mix

    Follow the published TRL `async_grpo_harbor` recipe: tasks in Harbor format served by OpenEnv, the harness run as shipped, two B200 GPUs per harness. The card's training curve and trajectories show what to expect.

Before you commit

Setup
Install pi with `npm install -g --ignore-scripts @earendil-works/pi-coding-agent` (Node.js 22.19+), serve Qwen3.8-27B with the LoRA in vLLM, and add a local provider with `"api": "openai-completions"` to ~/.pi/agent/models.json.
Integrations
vLLM (LoRA or merged), transformers serve, llama.cpp and Llama.app via the GGUF adapter, the pi-llama extension; sibling adapters exist for Hermes Agent and OpenCode.
Data and privacy
Self-hosted, so code and prompts stay on your machine. Training used public Harbor task sets and HF sandboxes with tool calls screened by safeagents; evaluation sandboxes had internet access, and the card found no lookups of upstream fixes.
Cost at scale
No licence fee; the cost is serving a 27B model. Per the card, the base model with thinking used a median 48.8k generated tokens per SWE-bench Pro task against 9.4k for the adapter, so the same GPU handles several times more tasks.
Limits
Thinking off is the trained setting; 100-step, 131,072-token evaluation budget; only works on Qwen3.8-27B; vision tower not served; prefix caching in vLLM needs `--mamba-cache-mode align`.

What’s good

  • SWE-bench Pro hard subset from 39.2% to 64.7% under pi with thinking off
  • About 4.5 times fewer generated tokens than the base model thinking
  • Under half the agent turns of the base model on tasks both solve
  • Corrections, trajectories and training recipe all published
  • Apache 2.0; runs as a LoRA in vLLM or llama.cpp, or merged

What’s not

  • Evaluated only under the pi harness with the completions API
  • GGUF adapter not re-evaluated; low-bit base quants drift
  • No gain over the base model once thinking is on
  • Needs the full 27B base model; vision is not served

Best for

Developers running the pi coding agent on their own GPU or Mac who want better results with thinking off and shorter, cheaper sessions.

Qwen3.8-27B-pi-adapted alternatives

Qwen3.8-27B-pi-adapted: common questions

Is Qwen3.8-27B-pi-adapted worth it?
A small, well-documented adapter that makes Qwen3.8-27B a far better pi agent with thinking off, and a useful recipe even if you never run pi. We rate it 3.9 out of 5.
How much does Qwen3.8-27B-pi-adapted cost?
Free under Apache 2.0, like the Qwen3.8-27B base it requires. Running it means self-hosting the 27B base; a 4-bit GGUF of the base is about 17GB.
What are the downsides of Qwen3.8-27B-pi-adapted?
Evaluated only under the pi harness with the completions API; gGUF adapter not re-evaluated; low-bit base quants drift; no gain over the base model once thinking is on; needs the full 27B base model; vision is not served.
Who is Qwen3.8-27B-pi-adapted best for?
Developers running the pi coding agent on their own GPU or Mac who want better results with thinking off and shorter, cheaper sessions.
How much does Qwen3.8-27B-pi-adapted cost for a team or in production?
No licence fee; the cost is serving a 27B model. Per the card, the base model with thinking used a median 48.8k generated tokens per SWE-bench Pro task against 9.4k for the adapter, so the same GPU handles several times more tasks.
Is Qwen3.8-27B-pi-adapted safe for company data?
Self-hosted, so code and prompts stay on your machine. Training used public Harbor task sets and HF sandboxes with tool calls screened by safeagents; evaluation sandboxes had internet access, and the card found no lookups of upstream fixes.
What are the best Qwen3.8-27B-pi-adapted alternatives?
The closest alternatives we have reviewed are Qwen3.8-27B, Inkling 2 Mini and Gemma 4.

Answers checked Nov 21, 2026.