DeepSeek updates V4.1-Flash AI model with faster task execution

Chinese AI startup DeepSeek has released an updated version of its V4.1-Flash model, claiming improvements in task execution speed.

According to the model card on Hugging Face, the update focuses on reducing execution time and output length, though DeepSeek did not provide specific metrics for the improvements. The company’s internal testing showed modest gains on its software engineering benchmark, with the model now reportedly resolving a slightly higher percentage of tasks than its predecessor. However, comparisons with competing models like Anthropic’s Claude Opus remain mixed: while DeepSeek’s model leads in some benchmarks, it trails in others, including a notable gap on one key test.

Pricing continues to be a focal point for DeepSeek. The updated model maintains the same API rates as the September release, which were significantly lower than those of U.S.-based competitors. This cost advantage appears to have had an impact on the market, as shares of some Chinese AI firms reportedly declined following the original V4.1-Flash launch. The updated model is now available in DeepSeek’s web and mobile apps, with API access unchanged.

Alongside the model update, DeepSeek introduced new tools for developers building agent-based applications. The company opened its search engine to external users through an API that returns web results, along with a separate tool for converting web pages into structured formats. Both tools carry a per-request fee, though the company offered temporary free access for certain integrations. The update also appears tailored for agent software, including DeepSeek’s own tools and third-party products like Tencent’s coding assistant and desktop agent, which have supported V4.1-Flash since its initial release.

The timing of the update follows a recent service disruption. Earlier this month, DeepSeek’s API experienced an extended outage, which some observers attributed to infrastructure challenges—a recurring issue for Chinese AI firms navigating compute constraints. The company responded with credits for affected users and pledged to improve system reliability.

DeepSeek’s approach stands out for its emphasis on cost and iteration speed, rather than headline-grabbing benchmark performance. While the company’s open-source model weights (available under the MIT license) could pressure closed-source competitors, its reliance on self-reported metrics makes independent validation difficult. This strategy aligns with broader trends in China’s AI sector, where firms often prioritize rapid deployment and affordability over incremental performance gains.

The update arrives as DeepSeek pursues a significant funding round, reportedly aiming to double its initial target. If successful, the raise would further solidify its position in China’s AI landscape, though questions remain about its ability to scale infrastructure and sustain its pricing advantage. For now, the V4.1-Flash update serves as another example of how Chinese AI firms are competing not just on cost, but on the speed of iteration—a race where small, frequent improvements may prove as consequential as major breakthroughs.

Sources: siliconangle.com