ZeroEntropy merges into Notion: knowledge work tools begin to develop their own models

Notion announced a seemingly low-key but meaningful move on July 1: the ZeroEntropy team joined the company. The startup specializes in "efficient, task-specific models for knowledge work" - pursuing a balance of efficiency and effectiveness on specific tasks. For Notion, this is a key enhancement from "calling other people's models" to "self-developed model layer".

Why is it a "knowledge work-specific model"

ZeroEntropy's entry point is very precise: general-purpose large models are often "killing a chicken with a knife" in knowledge work scenarios - redundant capabilities, high costs, and unstable performance on specific tasks. The "task-specific model" pursues specific tasks such as document sorting, knowledge retrieval, and information extraction, using smaller models to achieve more stability and efficiency. This is exactly the underlying capability that a knowledge management tool like Notion needs most.

A complete AI strategy puzzle

Putting together Notion's actions in the first half of 2026, the intention is clear: intensive release of AI capabilities from May to June - Notion Developer Platform (expanding the building blocks of Notion and Agent), Custom Agents (AI colleagues that can autonomously handle complete workflows, has ended public beta), Notion AI and Multi-region Data Architecture (ensuring that customer data is processed and stored in the region where it is located). The addition of ZeroEntropy just completes the "model self-research" capability at the bottom of this puzzle, reducing dependence on external model providers.

From an industry perspective, Notion strengthens its positioning as a "knowledge work operating system + AI" through "acquisition of model teams", reflecting a trend: application layer companies begin to penetrate into the upstream model layer. It has formed a cooperative and competitive relationship with basic model manufacturers such as OpenAI and Anthropic - using other people's general models as the basis and its own special models for differentiation. For the domestic knowledge management/collaborative office track (Feishu, DingTalk, Yuque, etc.), this prompts a strategic proposition: should office tools in the AI ​​era also consider a "self-developed model layer" to build a moat, rather than permanently renting the capabilities of others?

Several directions worth tracking in the future:

1. The specific implementation of the ZeroEntropy model: Which Notion function will be embedded first (search, sorting or Agent).
2. The impact of self-developed models on the cost structure: Can knowledge work-specific models significantly reduce Notion’s reasoning costs.
3. Evolution of relationship with OpenAI/Anthropic: Will self-developed models squeeze the share of external models in Notion.
4. Decision on self-development of domestic office tools: Will Feishu and DingTalk follow the "self-developed special model" route.