DeepSeek V4.1 Flash Improves Software Tasks Through Self-Feedback
DeepSeek’s V4.1 Flash 1017, released about five weeks after its predecessor without a new version number, shows marked gains in software generation and computer use, according to Two Minute Papers host Károly Zsolnai-Fehér. He points to the release’s benchmark results and demos of games and 3D scenes, and says the update’s use of reinforcement learning with self-feedback may help explain the improvement despite no architectural change. Its low token price, he cautions, does not translate directly into the same cost advantage per task.
The capability jump is larger than the version number suggests
DeepSeek V4.1 Flash 1017 is presented as a practical leap in building software: examples include a playable mobile game, interactive 3D scenes and detailed 3D models. The update arrived only about five weeks after V4.1 Flash, without a new version number. For Károly Zsolnai-Fehér that modest label raises the central question: what changed if the architecture did not?
A benchmark graphic compares the versions on ProgramBench and Terminal-Bench 4.0, highlighting gains of 8.8 and 11.1 points, respectively. Zsolnai-Fehér says the model is substantially better at producing working software and using a terminal to operate a computer. DeepSeek’s release notes make a related claim: stronger benchmark performance, faster task completion and shorter outputs, describing the update as a performance-and-cost improvement.
The demonstrations give that claim a concrete shape. One shows a mobile game with a bird character moving through a forest level; another shows an interactive koi pond, where the interface invites users to drag the jelly-like scene and tap for a ripple. Other examples include a rendered city and a jet engine model. The benchmark comparison and examples frame the change as especially relevant to end-to-end project work.
The price per token is not the price per task
DeepSeek’s API price is a little more than a dollar per million output tokens, which Zsolnai-Fehér compares with Claude Opus 5.5. On that measure, the DeepSeek model is 15 times cheaper. But he cautions that dollars per million tokens do not tell you what it costs to complete a particular job.
In his independent tests, DeepSeek can use two to three times as many tokens as Claude Opus to accomplish the same task. That reduces the headline advantage when the comparison is made per task, though Zsolnai-Fehér still considers DeepSeek’s task cost good—just “a bit less good” than the advertised per-token comparison implies.
The relevant comparison is the cost of completing the same task, not simply the listed cost of a fixed number of tokens. Models differ in how many tokens they use to finish a request, so a low token price alone does not determine which system is cheaper for a given job.
The weights, meanwhile, are available to download for free. Running them locally is another matter: Zsolnai-Fehér jokes that the hardware would cost “exactly one bazillion dollars.” His point is not that an individual can necessarily run the model at home. It is that open weights can fuel a chain of competition: one system approaches frontier capability, then others try to make comparable systems smaller and cheaper.
Self-feedback may help explain what changed without a new architecture
Zsolnai-Fehér says the architecture did not change, sharpening the question of how capabilities could improve so much in about five weeks. He examines DeepSeek’s release announcement, which says V4.1 Flash 1017 uses reinforcement learning with self-feedback to improve end-to-end project development, including game creation and 3D modeling. For him, “learning from self-feedback” is the key clue.
This means that the AI is able to critique its own work, and learn from it.
The idea differs from relying only on more human-provided training data. In the self-rewarding approach Zsolnai-Fehér describes, a language model is prompted to act as a judge and provide rewards during training. A paper shown on screen, Self-Rewarding Language Models, reports that iterative training improved both instruction-following and the model’s ability to provide high-quality rewards to itself.
A racing game illustrates how such a feedback loop might work. The model writes a game and plays it, notices that collisions do not work, revises the code and tries again. It might then identify missing acceleration or weak visual detail and improve those too. Each attempt supplies experience that can guide the next one, rather than requiring a person to specify every defect in advance.
The hard part is evaluation. Zsolnai-Fehér notes that the model has to judge whether its own output is good, and that “it’s not always easy to tell when it is doing well or not.” He presents self-feedback as a possible explanation for the gains, not a settled account of every improvement.
Open weights extend the competition, but not away from hardware costs
For Zsolnai-Fehér, the broader significance is a relay: a new AI system appears, then others try to build something nearly as capable at a fraction of its size and cost. DeepSeek, in his framing, pushes toward frontier models; other systems can then distill DeepSeek into smaller, cheaper ones.
That does not make the current model inexpensive to run locally. Zsolnai-Fehér says he cannot run it at home, and suggests that a system needing costly hardware now might be runnable on a powerful laptop less than a year later. The practical implication is a possibility, not an immediate promise: open weights make the model available to build on, while deployment still depends on access to substantial compute.