← Back to feed

Lenovo TianxiAI announced on October 9 that its TianxiCode agent framework achieved a 71% solve rate on SWE-bench-Live Lite, earning the top global ranking with official Verified certification. The foundation model underneath TianxiCode is DeepSeek V4.1 Flash.

That last detail matters. DeepSeek V4.1 Flash is an open-weight model that scores 79.0% on the standard SWE-bench Verified mini-SWE-agent harness — well below the 95% territory of Fable 5.1 and Claude Opus 5. TianxiCode extracts a top-ranked result on a harder, contamination-resistant benchmark using a model that isn’t in the top tier of foundation capability.

What SWE-bench-Live Measures

SWE-bench-Live uses real GitHub issues filed after model training cutoffs. There is no data leakage possible because the issues postdate the model’s knowledge. That makes it harder and more credible than the standard verified dataset, where contamination from training data is a live concern.

A 71% solve rate on SWE-bench-Live Lite with Verified certification means a third party confirmed the evaluation setup and the results. It is not vendor-reported only.

The Agent Architecture

TianxiCode’s stated capabilities:

• Cross-file multi-hop retrieval for navigating large codebases
• Autonomous planning with multi-turn tool calling
• Closed-loop patch generation with test-driven self-correction

The self-correction loop is where most of the gap between foundation model capability and agent capability gets created. A model that can generate a patch, run tests, observe the failure, and regenerate a corrected patch closes much of the distance between an 80% foundation model and a 90%+ system-level result.

Lenovo describes DeepSeek V4.1 Flash as the “reasoning engine” and TianxiCode as the “engineering workflow” — a clean separation between what the model knows and how the agent applies it.

Hardware Deployment Plans

Lenovo says TianxiCode will be integrated into its TianxiAI hardware product lineup. The stated goal is reducing developer debugging and engineering coordination costs by embedding autonomous code-repair tools directly into device-level toolchains.

This is a distinct commercialization model from API-based coding agents. If TianxiCode ships as part of ThinkPad or workstation firmware-level developer tools, the distribution channel is hardware sales rather than per-call inference pricing.

The Foundation Model Cost

DeepSeek V4.1 Flash is available at approximately $0.029 per task on LiveBench’s cost-per-successful-task metric. That is the cheapest per-task cost of any model in the top 10 on LiveBench’s agentic coding subcategory. A code agent running TianxiCode on DeepSeek V4.1 Flash runs at a fraction of the cost of the same framework on Claude Fable 5.1 ($1.212/task) or Claude Opus 5.5 ($0.799/task).

The benchmark result and the cost structure together make TianxiCode a credible candidate for high-volume automated code review pipelines, where per-call cost matters and contamination-resistant benchmarks provide the performance guarantee.