DeepSeek 版本命名规则解析：V4.1 与 V4 几乎无关

DeepSeek has great discipline in model nomenclature: leading number = current base model, decimals = nontrivially altered checkpoints (model merge, continued pretraining+quantization, attention adaptation), MMDD = a routine post-training update. There are only two exceptions, which are still throwing everyone off: R1 ("R2 wen?") and now V4.1 which is not related to V4 models whatsoever.

There's a mundane, deflationary explanation for both cases. R1 was a performative response to o1, so it couldn't be called V3-Reasoner; by V3.1, they got rid of the separate R-line, as they did for Vx-Coder before that, and then for separate Vx-Provers and Math and Vision. 4-Pro was a failure and they are too early in the release cycle to admit that they're salvaged its pieces for V5 (and released Flash first because compute-poor). You can think like this.

But I guess their method holds in spirit. They don't believe V4.1 is a big enough update. It's just a V4 that's made to not suck, and it so happens that V4 couldn't have been improved without pretraining from scratch. They are a spiritual sect and a lab doing basic research towards AGI before they're a commercial company shipping products. These "Versions" correspond to a significant question of ML research that they're trying to figure out. Admittedly, the questions are not purely scientific, because AGI is a heavy industry and not just clever math, so there is a certain dialectical tick-tock profile to their progress, quantity enabling new quality. The next question is usually pre-announced in the conclusion of a given mainline paper. V1 was an investigation into basic data, model and compute scaling laws for LLMs, and they announced then that the next targets will include MoEs and multimodality. Along the way they tackled data collection and RL. V2 was a study in compression and sparsity which permit greater scaling. V3 era, if you fold Rs into it, was mainly about RL for reasoning, bolted on top of infrastructural exploration that built on V2. Then, they started thinking about LLMs in the limit; V3 paper announced "infinite context" as the goal, and much of 2025 was spent on research into longer contexts and infrastructure that will permit them in practice. NSA, DSA, and eventual V4, and then V4.1 that delivers on its promise (or so it seems). 1M contexts are now a no-brainer for Whale. With RLM and modern harnesses, this means no limit (or rather, no limit within the human range of ability) on the scope of project that is accessible for the model's effective working memory (especially with residual stream multiplexing). But Liang believes that working memory won't be enough. V5 has to answer a set of questions too. I guess they want an MVP of architecture that enables continual learning. This may take another year or more.