Short verdict: The finest moment of spring.
Basic situation:
What sets DeepSeek apart is that they are passionate about—and good at—improving efficiency through underlying innovation. Through such unremitting effort, even if they take some detours, they ultimately achieve breakthrough results. This time, V4.1 appears to be a minor version, but from model architecture to parameters, many major changes have taken place.
Among these, its logical reasoning capability has risen significantly compared with its own Pro, reaching the head of the first tier among domestic models. The extreme speed brought by underlying architecture innovation also offsets part of its disadvantage in thinking length, giving it the lowest average latency in its tier. Moreover, if the high setting is used, thinking length drops markedly and response latency falls sharply, entering a highly usable range. Even globally, it is competitive.
Agent Coding capability, however, improved only to a limited extent, slightly better than the previous-generation Pro. But V4.1's Agent behavior has changed considerably compared with before, with large differences across tasks; this will be elaborated below.
Logic scores:
The table's sorting has been switched to descending by median score.
*1 The table highlights comparison relationships, so it shows only some comparable models, not the complete ranking.
*2 For questions and testing methods, see: Large Language Model Logic Ability Benchmark – 26-08 Monthly Leaderboard, with #76, #77, #78 added.
*3 The full leaderboard is updated at https://llm2014.github.io/llm_benchmark/
Agent scores:
Coding tasks: V4.1 Flash is overall better than versions such as V4 Pro and V4 Flash Vision. Looking at test scores alone cannot fully reflect this. Specifically, V4.1 Flash's overall aesthetic ability is significantly higher than all V4 versions, but it still lags behind the first tier. V4.1 Flash can use visual capabilities to inspect and correct many UI presentations and some layout and style issues, but for tasks with higher aesthetic demands—such as spacing, color schemes, and more complex advanced modeling, motion effects, and transitions—V4.1 Flash performs below expectations. V4.1 Flash's repair ability is strong; for most aesthetic problems, if there is human intervention or even a provided template, the model can do better.
V4.1 Flash takes significantly more steps to complete tasks than the previous-generation Pro and Flash. The observed increase is at least +20%, and in an extreme case, Project J, the increase was +700%. Most of this increase comes from self-testing. The previous-generation Pro self-tested more than the previous-generation Flash, but overall it remained in a reasonable range; on less familiar tech stacks, it basically did not self-test at all, ending after only simple static compilation checks. V4.1 Flash, by contrast, conducts more detailed, comprehensive, and extensive self-testing, writing test cases for each feature point and verifying screenshots. This kind of behavior in max mode usually comes from reinforcement learning environments, where the model makes cost-no-object attempts to chase a very small reward gain. If the high or low setting is used, performance on medium-difficulty tasks is similar to max mode, while the number of steps is much lower, making it more practical.
V4.1 Flash's bug localization ability is comparable to the previous-generation Pro and far better than the previous-generation Flash. It can find root causes and correctly fix any bug without human reminders. The difference from Pro is only that for a few bugs with multiple possible fixes, V4.1 initially corrected in the wrong direction. But this is also random. Moreover, this is limited to test tasks; for broader real-world tasks, readers will need to observe for themselves.
Inference consumption: V4.1 Flash in max mode shows an unprecedented level of reasoning effort. On multiple medium-difficulty tasks, it can even consume 260–340K, far higher than the previous generation and all other models. In simple instruction-following tasks, it also consumes at the 60K level. More critically, after V4.1 Flash consumes excessive tokens, its problem-solving accuracy does not improve; instead, the previous-generation Pro has a higher average accuracy. On tasks where V4.1 Flash actually scores higher, consumption is basically on par with the previous generation. This once again confirms that max mode is not suitable for daily use.
Hallucination performance: V4.1 Flash hallucinates less than the previous generation. However, the previous generation's hallucination control was not excellent, and in some scenarios it was below models in the same tier. V4.1 Flash can be considered to have filled this shortcoming, but compared with internationally leading models, the gap remains obvious. The biggest problem is that once the text itself becomes more confusing, V4.1 Flash still cannot handle it, performing very unstably; in the worst case, it is no better than the previous generation. In Agent tasks, this shows up as possibly randomly dropping some information when facing a complex task list.
The Cyber Historian says:
The development of large models is spiral. The success of the previous generation does not mean the successor model will comprehensively surpass it, and the shortcomings of the previous generation do not mean the team has exhausted its talent and run out of ideas. The key is what direction the team insists on, and whether their actions are on the main path. For DeepSeek, which pursues AI for all, extreme cost reduction is precisely its persistent beacon. And every innovation from DeepSeek brings new possibilities to the entire model world, ultimately feeding back into progress across the industry. In this regard, it feels like a warm spring breeze.