‘Faster and leaner’: DeepSeek rolls out upgraded AI model after recent service disruptions

DeepSeek's new model, DeepSeek-V4.1-Flash-1017, shows reported gains in several benchmark tests. The company claims its score on Humanity’s Last Exam without tools increased from 36.8 to 42.6, while Terminal-Bench 4.0 for coding agents rose from 31.2 to 42.3, and DeepSWE v1.1 improved from 74.2 to 77.0. These figures, published on Hugging Face and in DeepSeek's technical report, indicate progress in difficult questions across various domains and in AI agent tasks.

The release follows recent service disruptions, including slowdowns and a five-hour outage earlier in the week, which some experts attributed to a shortage of computing resources. This suggests that high demand for DeepSeek's competitively priced AI models may have strained its existing infrastructure. The company's stated goal of reducing response times and producing shorter outputs with the new model could be a direct response to these operational challenges.

The reported improvements in speed and efficiency, alongside the potential for lower operational costs, position DeepSeek to remain competitive in the rapidly evolving AI landscape. Chinese AI developers, including DeepSeek, are actively vying with both US firms like OpenAI and Anthropic, and domestic rivals, to deliver more capable systems at reduced costs. This dynamic environment encourages continuous iteration and optimization of AI models.