The Tianxi AI team from Lenovo has developed a coding agent that secured first place in the SWE-bench-Live platform's Lite ranking. This benchmark requires artificial intelligence systems to resolve real issues taken from GitHub projects. The TianxiCode agent, which utilizes DeepSeek-V4.1-Flash as its base model, managed to solve 71% of the tasks and passed the official benchmark verification, according to a statement from Lenovo distributed by QbitAI on October 9th.
It is important to note that this result pertains exclusively to the Lite section, not the full or multilingual rating tables. According to the public leaderboard, the entry from October 8th showed the resolution of 213 out of 300 Lite tasks. The advantage over the next verified participant in the same category is only 0.67 percentage points, as the next participant is at 70.33%.
SWE-bench-Live differs from traditional coding tests that check syntax or individual functions. Each task is based on a real GitHub ticket, and the system must create a patch that fixes the issue in a reproducible execution environment. To receive the 'Verified' tag, teams provide the complete workflow trajectory of their agent, after which organizers check the evaluation input data and isolation setup for leakage of reference answers, test examples, or results.
How the TianxiCode Agent Works
Lenovo describes the function distribution as follows: DeepSeek-V4.1-Flash acts as the 'engineer's brain,' analyzing code and suggesting fixes, while TianxiCode provides the 'eyes, hands, tools, and work discipline.' The company emphasizes three key capabilities. Multi-step information retrieval across files, combined with context pruning and slicing, allows the agent to trace the root cause of an error in a large codebase. Autonomous planning with multi-turn tool calls enables it to formulate a debugging plan, run tests in the terminal, read logs, compare change history, and then adjust its approach if difficulties arise.
A closed-loop correction cycle with self-correction based on tests allows new code to be run in an isolated environment and continued refinement until the patch passes verification. TianxiCode is a component for code generation and software development within Lenovo's Tianxi AI ecosystem. Lenovo plans to integrate this work into developer toolchains and its AI-based hardware products, although timelines or specific devices for its operation have not been announced.
The achieved result also demonstrates the potential of an economically efficient Flash-class model when combined with a powerful agent mechanism. Lenovo has not published separate evaluations of the same system on other models.