Deepseek's latest AI model is smarter than ever, but its low running costs come at a price

Deepseek has unveiled an upgraded version of its open-weight AI model, V4.1 Flash, now at revision 1017. While the model boasts remarkable efficiency and cost-effectiveness, it is not without its flaws. Deepseek positions the model as a solution for open-ended tasks with broad descriptions, suitable for LLMs that fall between GPT-5.6 Sol and Claude Opus 5 on internal benchmarks.

However, the model exhibits a tendency toward hallucinations, a common issue in AI models that struggle to recognize when they lack supporting facts. In a benchmark test, the model answered correctly 43% of the time, whereas competitors managed scores up to 58. Moreover, the model hallucinated responses 67% of the time for non-answers.

Despite these shortcomings, Deepseek's V4.1 Flash remains a powerful and affordable option for specific tasks that require verification, such as programming, OCR with review, and format conversions. However, users should exercise caution when employing the model for open-ended research, fact-checking, or interpreting precise documents, as it is likely to produce unreliable results due to its propensity for hallucinations.