Tag
A LoRA adapter for Qwen3.8-27B, trained with reinforcement learning to power the pi coding agent, tops the Hugging Face hub trends — matching the base model's xhigh thinking accuracy on SWE-bench Pro and Terminal-Bench 2.1 with about 4.5× fewer tokens.
A LoRA adapter for Qwen3.8-27B, trained with reinforcement learning to serve as the model behind the OpenCode coding agent, matching base-model thinking-mode accuracy on SWE-bench Pro and Terminal-Bench while using 6-8x fewer tokens.
A LoRA adapter for Qwen3.8-27B trained with reinforcement learning to power the pi coding agent, matching the base model's highest-effort thinking accuracy on SWE-bench Pro and Terminal-Bench while generating about 4.5x fewer tokens with thinking off.
Lev is a 4B open-source AI model based on Qwen, designed for efficient typed decision-making in a single forward pass with zero output tokens, achieving strong benchmarks for its size.
The paper identifies a blind spot in machine unlearning benchmarks: underrepresentation of causal (Why-type) knowledge, and proposes 5WBench, a balanced benchmark, and Maat, a three-phase unlearning framework on LoRA adapters that achieves high forgetting and retention on causal facts.