Meituan LongCat open-sources LongCat-DeepResearch research harness
Meituan LongCat released the LongCat-DeepResearch repository, an open-source harness for open-ended research. It is MIT-licensed and requires Python 3.10 or newer. The repository includes only the research harness; data-construction pipeline, datasets, rubrics, trajectories, and training code are not included.
China context
- Original name
- 美团龙猫
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate the harness with their own LLM, web search, and web fetch backends; the MIT license allows commercial use, but they must supply their own model and services.
- For investors
- Meituan LongCat's open-source release may pressure proprietary deep-research offerings, but the lack of released model weights or training data limits immediate competitive impact.
Meituan LongCat open-sourced LongCat-DeepResearch, a research harness that captures evolving research requirements in an executable ResearchSpec, supports independent section research and writing, and makes targeted revisions to assembled reports. Users can run deep-research tasks with their own services by implementing llm, web search, and web fetch according to the backend contract. Paired with a LongCat-2.0-based model enhanced for deep research, the complete system outperforms Deep Research offerings from ChatGPT, Claude, and Gemini on DeepResearchBench, DeepResearchBench II, and ResearchRubrics. The repository is MIT-licensed and supports Python 3.10 or newer.
The harness uses a three-stage pipeline: explore and plan (multiple planners, judge, critic, reviser produce an executable ResearchSpec), research sections (independent researchers in separate contexts), and assemble and edit (global editor identifies ownership/consistency issues, local editors apply section-scoped revisions). The backend contract defines llm, web search, and web fetch methods; the harness passes web search and web fetch as OpenAI-style function tools in llm requests. The repository includes only the harness, not the data-construction pipeline, datasets, hidden rubrics, training trajectories, or training code.
Developers outside China can now integrate a competitive deep-research harness into their own stacks without relying on hosted services, reducing dependency on proprietary offerings from OpenAI, Anthropic, and Google. The MIT license permits commercial use, but the absence of the data-construction pipeline and training code limits replication of the full system's performance.
The open-source harness lowers the barrier for enterprises and developers to build custom deep-research agents with their own models and search/fetch services, reducing costs compared to subscription-based offerings. However, the value depends on the quality of the user's backend implementation and the availability of a suitable model.
Observable next signals include whether Meituan LongCat releases the LongCat-2.0-based model weights or API, publishes the data-construction pipeline or datasets, or provides hosted backend examples. Independent verification of the benchmark claims would require access to DeepResearchBench, DeepResearchBench II, and ResearchRubrics, which are not included in the repository.