AIBRIEF
On July 31, 2026 DeepSeek released V4 Flash 0731 for public API testing, keeping the same base model and weights—NetEase Youdao calls it a "software tune up" of an existing model, not a new one—as it shipped LobsterAI, NetEase Youdao's new
On July 31, 2026, DeepSeek opened the V4-Flash-0731 checkpoint to public API testing and disclosed in its Change Log that the new release "keeps the same model architecture and size" as the May V4-Flash-Preview and "was only re-post-trained." The model card is on Hugging Face.
Within days, NetEase Youdao moved its dictionary, translation, simultaneous interpretation, language-practice, and desktop-assistant products onto the new checkpoint, per a Leiphone write-up of the integration. The lead product is LobsterAI, which Youdao calls the first 100% open-source desktop-grade office Agent from a major Chinese tech vendor. The same write-up cites 5,700+ GitHub stars and "hundreds of thousands of daily interactions", both company-supplied numbers pending independent verification.
Youdao calls the change "底座没变，变的是后天教育": the base didn't change, only the "second education" (post-training) did. That bound comes from DeepSeek itself: any product lift comes from better tool calling, multi-step task execution, and long-context handling, not new weights.
LobsterAI ships three new capabilities: a role-based "Expert Kit," one-click website building, and multilingual voice input. Youdao's ThinkFlow enterprise platform, which already aggregates more than 20 LLMs, has also moved to V4-Flash-0731.
What remains unknown: whether the post-training-only lever delivers measurable product-grade gains once LobsterAI is tested outside NetEase Youdao, and whether the LobsterAI adoption numbers hold up against the public GitHub repo.