China Tightens AI Export Controls as DeepSeek Releases Improved Model

What Actually Changed

China plans to implement export controls on 23 key AI technologies, including model weights, source code, and training methods for systems exceeding 10^26 FLOPs or 2 trillion parameters. The restrictions will likely impact Western companies using Chinese AI technologies. Separately, DeepSeek released V4.1-Flash-1017, claiming it as a Pareto improvement over its September release with stronger scores, faster task completion, and shorter outputs. The company also opened its Search API alongside this release. A new benchmark called Taste-Bench revealed that even the best model tested, GPT-5.6 Sol, correctly answers only under 60% of simple binary questions in long-running tasks.

Why It Matters

China's export controls represent a significant geopolitical shift in AI technology distribution, potentially fragmenting the global AI ecosystem. The restrictions may push Western companies to double down on US open-weight models over Chinese alternatives, creating a bifurcated market. DeepSeek's improved model demonstrates the ongoing competition in optimizing performance and cost, particularly important as enterprises evaluate deployment options. The Taste-Bench results highlight a fundamental limitation in current AI agents' strategic decision-making capabilities, suggesting that despite advances in reasoning, long-term planning remains a significant challenge for foundation models.

Capability and Benchmarks

The new Taste-Benchmark specifically tests AI agents' ability to make strategic choices during long-running tasks, revealing that even the best-performing model, GPT-5.6 Sol, achieves less than 60% accuracy on simple binary questions. Notably, the research found that allowing models to think longer does not improve performance in these scenarios. Meanwhile, DeepSeek's V4.1-Flash-1017 achieved a score of 77.0 on DeepSWE, demonstrating strong capabilities in software engineering tasks. The company claims this latest version represents a Pareto improvement in both performance and cost efficiency over previous releases.

What We Are Watching

We're monitoring how China's export controls will impact the global distribution of AI technologies and whether they will accelerate the adoption of US open-weight models as predicted. The implementation timeline and specific enforcement mechanisms of these controls remain unclear. We're also tracking the response from Western companies to these restrictions and potential countermeasures. Additionally, we're watching how the industry addresses the demonstrated limitations in AI agents' long-term decision-making capabilities, as identified by Taste-Bench, and whether this will drive new research directions or model architectures.