(Yicai) Sept. 9 -- Chinese artificial intelligence model developer DeepSeek has begun internal beta testing of the V4.1 Flash intermediate version, with early testers reporting that the model balances multimodality and speed and is "incredibly fast."

V4.1 Flash adopts a new model architecture, providing native multimodal support, enhanced capabilities, faster speeds, and lower costs, DeepSeek said yesterday. The initial tests implement a concurrency limit of 20 requests per account.

The internal beta testing will expire and go offline tomorrow, meaning developers only have two days of access.

"The first impression is that it's incredibly fast," some testers pointed out. The new V4.1 Flash significantly outperforms V4 Flash Vision-Exp in end-to-end processing for the same task, with scalable vector graphics code generation six times faster and long-context retrieval more than five times faster, they noted.

The V4 Flash update is not just fine-tuning of the previous version but involves a complete retraining of the model. The model's new architecture is expected to deliver a significant leap in capabilities, potentially propelling DeepSeek back into the top tier of global model performance.

However, regarding the "lower costs," some testers said "CNY10 (USD1.49) gets you only five minutes," while others pointed out that "tens of Chinese yuan can disappear in an instant." This has led to speculation that the increased speed may also lead to faster token consumption over the same time.

The price in the beta testing is the same as for DeepSeek-V4-Flash, so based on the per-token price developers pay, the new model has not been discounted. What DeepSeek calls "lower costs" may imply improved inference efficiency or reduced computational cost and token usage required for the same tasks.

DeepSeek's updates and iterations have noticeably accelerated. In less than 40 days, the company has continuously updated its model, native multimodal capabilities, and agent development tools, indicating that it is working to complete its product system from the underlying models to the agent infrastructure.

However, DeepSeek's model capabilities ranking has dropped this year, with the firm failing to secure a leading position in the latest round of model competition. Its model achieved an intelligence index score of 36, ranking 15th worldwide, according to AI benchmarking organization Artificial Analysis.

Claude and GPT -6 hold the top two positions on the list. Among Chinese models, Zhipu AI ranks seventh, Kimi ninth, and Qwen 13th.

Editor: Martin Kadiev