Qwen3.8-27B reportedly jumped from 51.7% to 86.52% on Terminal-Bench 2.1 after Tencent Hunyuan applied a recursive self-improvement method. That is a gain of 34.82 percentage points without increasing the model’s parameter count. The method, called Terminal-RSI, focuses on a problem that matters for AI agents: completing tasks reliably, not just generating commands that look correct. A terminal agent may need to inspect a codebase, modify files, install dependencies, run tests, recover from errors, and verify that everything works. Getting most of these steps right is not enough if the final task remains incomplete. Terminal-RSI tackles this by turning successful executions into training data and repeatedly refining the model. How Terminal-RSI works The process alternates between two training stages, repeating the cycle twice. 1. Recursive Self-Rewrite (RSR) The model takes execution histories from verified successful runs and rewrites them into examples for supervised fine-tuning (SFT). This lets the model learn from successful behavior instead of treating each completed task as an isolated result. 2. Policy Refinement (RPR) Reinforcement learning then improves the model’s action policy. Critic information from the preceding stage is carried forward to help guide the next round of optimization. The complete cycle looks like this: RSR-1 → RPR-1 → RSR-2 → RPR-2 Each iteration uses the improved model to generate another round of training examples, followed by further policy refinement. The aim is to make successful task-completion behavior more consistent over time. The reported benchmark results Tencent Hunyuan reports improvements across several terminal-agent evaluations: Terminal-Bench 2.1: 51.70% → 86.52% Terminal-Bench 4.0: 0% → 19.84% Terminal-Bench Hard: 79% Long-horizon process reward: 0.21 → 0.37 Strict long-horizon completion: 0/46 → 5/46 SWR100: Reported success rate doubled The paper also reports average gains of roughly 30 percentage points across multiple benchmarks. These metrics measure different things, so they should not be treated as directly interchangeable. The long-horizon results are especially revealing. Strict completion improved from zero successful runs to five out of 46, but most attempts still failed to meet the full completion criterion. That distinction matters because making progress on a task is not the same as finishing it successfully. Why the training approach is interesting Traditional supervised fine-tuning teaches a model from prepared examples. Reinforcement learning adjusts behavior using feedback about its actions and outcomes. Terminal-RSI combines these approaches in a repeating loop, with several techniques supporting the process: Verified trajectory rewriting converts successful execution histories into new training examples. Critic weight inheritance carries useful optimization information between refinement stages. TB-GAE helps assign learning signals across sequences of actions. Success-Prioritized Reward places greater emphasis on completing tasks successfully rather than rewarding partial progress as if it were full completion. Together, these methods target the gap between issuing plausible commands and reliably achieving the intended result. What this means for smaller models Qwen3.8-27B shows why parameter count alone does not determine how useful an AI agent can be. A model’s performance also depends on its training, its ability to learn from feedback, and how effectively it handles multi-step tasks. Recursive self-improvement offers a way to build on existing capabilities without making the underlying model larger. The results are promising, but they remain benchmark-specific. Independent replication, evaluation across more environments, and testing on unfamiliar tasks will help establish how well the gains generalize.