Qwen3.8-27B-pi
A community fine-tune of Qwen3.8-27B that fixes effort-ordered reasoning so higher reasoning effort reliably costs more tokens and does better, not worse.
This entry is marked for review. Check the original sources before relying on its details.
Overview
Qwen3.8-27B-pi, published on Hugging Face as bytkim/Qwen3.8-27B-pi, is a community fine-tune of Qwen3.8-27B built for the Pi agent coding harness. The base Qwen3.8-27B model violated 'effort ordering': its low reasoning setting used roughly 1.37x more tokens than medium while passing fewer tasks, so raising the reasoning-effort slider could paradoxically lower quality and raise cost. The pi fine-tune runs supervised fine-tuning followed by GRPO reinforcement learning with a 'Success-Conditioned, Effort-Ordered Reward': failed rollouts score zero, xhigh gets no length penalty, and medium/low are scored against the token budget of the lowest higher-effort level that solved the task. On Terminal-Bench 2.1, Qwen3.8-27B-pi at medium reasoning hits 75.28% pass rate, matching the base model's xhigh pass rate while using roughly 41% fewer output tokens.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Fixes 'effort ordering' violations where low reasoning cost more tokens than medium in the base model
- Trained via SFT followed by GRPO with a Success-Conditioned, Effort-Ordered Reward
- Matches base model's xhigh pass rate on Terminal-Bench 2.1 at medium reasoning effort
- ~41% fewer output tokens than the base model for equivalent pass rate
- Tailored for the Pi agent coding harness's read-edit-run-respond loop
Use cases
- Agentic coding workloads where reasoning-effort/cost tradeoffs need to be predictable
- Running frontier-class coding agent behavior on a single consumer-class GPU
- Reducing inference token spend in multi-step coding agent loops
Pricing
Open weights on Hugging Face; free to download and self-host, compute cost depends on hardware used.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- Not listed
- Record updated
- 3 Oct 2026
- Record last checked
- 3 Oct 2026