LLMxMapReduce is an open-source framework from THUNLP, OpenBMB, and AI9Stars that helps large language models understand and generate very long sequences. Hosted on GitHub with about 880 stars under the Apache-2.0 license, it was still updated in 2026 and applies the classic MapReduce divide-and-conquer idea to LLM workloads.
Instead of feeding an entire long document to a model at once, LLMxMapReduce splits the input, processes the pieces in a structured way, and then combines the results. The first version introduced a structured information protocol and in-context confidence calibration to improve long-sequence understanding. Later versions extended the approach: V2, described in a 2025 arXiv paper, powers SurveyGO, an online survey-writing system, and V3 code was released in September 2025. The framework works with the OpenAI API and OpenAI-compatible APIs such as vLLM.
The repository collects the code for all three generations of the framework, so users can pick the version that matches their task. The first version demonstrated that even a compact model such as MiniCPM3-4B could handle long-sequence understanding when the work was decomposed properly, while the second generation generalized the approach and was applied to automated survey generation. Because it talks to standard OpenAI-style endpoints, teams can run the pipeline against hosted APIs or their own vLLM deployments without changing the overall design.
LLMxMapReduce is free and open source under the Apache-2.0 license. You only pay for the underlying model API calls you choose to make through your own provider accounts.
It is aimed at researchers and developers who need LLMs to read, summarize, or generate long documents reliably, especially where a single context window is not enough and a structured multi-stage pipeline gives better results.