{"id":370,"date":"2026-09-09T23:13:31","date_gmt":"2026-09-09T14:13:31","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/09\/09\/deepseek-v4-1-flash-rumors-and-features\/"},"modified":"2026-09-20T17:37:13","modified_gmt":"2026-09-20T08:37:13","slug":"deepseek-v4-1-flash-rumors-and-features","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/09\/deepseek-v4-1-flash-rumors-and-features\/","title":{"rendered":"DeepSeek V4.1 Flash: Faster and Cheaper Than V4 Pro"},"content":{"rendered":"

Overview<\/h2>\n

It has been revealed that DeepSeek plans to officially release its next-generation model, “V4.1 Flash”, around September 10, 2026 (Beijing time). According to reports, extensive internal and external testing shows that this new model comprehensively outperforms the existing “V4 Pro” across all key metrics of performance, cost, speed, and task completion time.<\/p>\n

Why It’s Making Noise<\/h2>\n

This news has gained significant attention on Hacker News, accumulating over 400 points and more than 200 comments. Particular interest has focused on the fact that a Flash model achieves performance exceeding a Pro model, alongside its overwhelming cost-performance ratio. Additionally, DeepSeek’s operational policy\u2014whereby the release of V4.1 Flash will route all requests for V4 Pro to V4.1 Flash at the Flash pricing tier\u2014has become a topic of discussion.<\/p>\n

Points of Discussion<\/h2>\n

The community is mainly exchanging views on the following three points:<\/p>\n

    \n
  1. \n

    Forced Model Switching and Operational Impact
    \nTechnical concerns have been raised regarding DeepSeek’s policy to redirect V4 Pro requests to V4.1 Flash and price them at the Flash rate. Pointing out that for users who have optimized their prompts for V4 Pro, unexpected model changes could undermine workflow stability.<\/p>\n<\/li>\n

  2. \n

    Language Adherence and Reasoning Settings Usability
    \nAs a characteristic of the Flash model, issues regarding language control have been noted. Reports indicate that language adherence is unstable, such as thinking processes or responses turning into Chinese even when asked in English. Furthermore, regarding the reasoning intensity settings, the lack of a practical “medium” setting among the three options (“low”, “high”, and “max”) makes controlling cost and time difficult, according to ergonomics feedback.<\/p>\n<\/li>\n

  3. \n

    Paradigm Shift Brought by Performance Improvements and Cost Reductions
    \nSurprise and expectations have been expressed over the emergence of a higher-performing and cheaper model, overturning the conventional wisdom that “Flash models are fast but inferior in performance.” Beyond simply “Flash replacing Pro,” attention is focused on the expansion of use cases where extremely low costs enable the automation of a vast number of small-scale, repetitive tasks that were previously unviable with large models.<\/p>\n<\/li>\n<\/ol>\n

    Community Reactions<\/h2>\n

    For each point of discussion, the following opinions have emerged from the community:<\/p>\n

    Forced Model Switching and Operational Impact<\/h3>\n

    Negative opinions are prominent regarding the forced model switching. Users who have already completed prompt testing and optimization for V4 Pro expressed concerns that even if the new model is “better,” models should not be swapped out arbitrarily on paying customers. Practical operational proposals have also been made, suggesting that the older model should be kept as “deprecated” for a certain period when introducing a new model to avoid sudden changes to validated workflow processes.<\/p>\n

    Language Adherence and Reasoning Settings Usability<\/h3>\n

    Regarding language control, practical issues have been reported. Pointing out that language adherence is unstable, such as the thinking chain and responses turning into Chinese when queried in English, or answers being returned in a language different from the input. Additionally, regarding reasoning intensity settings, opinions suggest that the current three levels (“low”, “high”, and “max”) lack a practical intermediate setting (“medium”). Voices are raising concerns that “low” is equivalent to turning reasoning off, while “high” and “max” take too much time and inflate costs, making it unergonomic.<\/p>\n

    Paradigm Shift Brought by Performance Improvements and Cost Reductions<\/h3>\n

    On the other hand, there are many highly positive reactions to DeepSeek’s evolution. Voices expressed astonishment at the pace at which the Flash model is overtaking the Pro model, alongside praise for its overwhelming cost-performance ratio. Specifically, the very cheap pricing structure is expected to expand the domain where a massive number of small, repetitive tasks\u2014previously cost-prohibitive with large models\u2014can be automated. Furthermore, some users reported specific hands-on experiences, noting superior comment-generation capabilities compared to other major models in certain coding tasks, as well as an astonishing speed of 300 to 400 tokens per second.<\/p>\n

    Sources<\/h2>\n