Lenovo's TianxiCode Agent and DeepSeek-V4.1-Flash Top SWE-bench-Live Lite at 71%, Verified

What happened

Lenovo's TianxiCode agent, using DeepSeek-V4.1-Flash, achieved 71% accuracy on the SWE-bench-Live Lite benchmark by generating reproducible code patches for real GitHub issues. The result was verified by the benchmark's organizers, who confirmed no leaks of test cases or answers.

Why it matters

The achievement highlights the potential of combining cost-efficient models with advanced agent frameworks to address complex coding challenges, offering a scalable approach for AI-driven software development.

• AI agents can effectively address real-world coding issues when paired with optimized models.
• Verification processes are critical for ensuring the integrity of AI benchmark results.
• Combining model efficiency with agent frameworks opens new possibilities for AI applications.

The bigger picture

The result demonstrates the viability of agent-based systems for code correction, potentially influencing future AI toolchains and hardware integration strategies.

Lenovo's success on SWE-bench-Live Lite underscores the growing capabilities of AI agents in software engineering, with implications for both tool development and hardware integration.

Technical details

• DeepSeek-V4.1-Flash: Causal encoder-decoder MoE model with 1M context and KV compression.
• TianxiCode's capabilities: Multi-hop file retrieval, context pruning, autonomous planning with tool calls, and closed-loop patching with self-correction.
• SWE-bench-Live Lite verification: Requires full agent run trajectories and isolation of evaluation environments.

FAQ