DeepSeek has commenced internal testing of its V4.1 Flash intermediate version, with the objective of assessing whether it can serve as a complete replacement for the current V4 Pro version. Key features of this new version encompass a revamped model architecture, built-in multimodal capabilities, and accelerated processing speeds. Concurrently, DeepSeek has announced a price reduction for its Flash series, effective from September 10th, in an effort to reduce costs for users.

To boost the intelligence density of reasoning chains, DeepSeek has integrated DSpark technology. This innovation enables an increase in overall service throughput while maintaining swift response times for individual users. For Agent scenarios, DeepSeek's open-source Harness tool streamlines execution processes, minimizes redundant calls, and eliminates unnecessary inputs, thereby cutting down on total task costs.

At present, several model companies are recalibrating their high-end capability strategies. For instance, Anthropic has rolled out Opus 5, which offers near-state-of-the-art capabilities at a more affordable price point. Should the Flash version prove capable of handling a substantial portion of the workload currently managed by the Pro version, the next-generation Pro version could then be tailored to tackle more intricate, long-term tasks.

Through optimizations in multimodality and execution frameworks, DeepSeek is striving for a holistic enhancement in efficiency, spanning from individual reasoning processes to entire tasks. Nevertheless, the full extent of these advancements remains subject to further validation.