Meituan releases MineExplorer, a long-horizon multimodal benchmark

On October 3, 2026, Meituan's technical team announced that the LongCat team at Meituan has built MineExplorer, which it calls the first benchmark for minute-level long-horizon tasks in open-world settings. According to the announcement, the benchmark is meant to systematically evaluate the real capabilities of multimodal large models on tasks that require long-horizon planning and that come with hidden prerequisites — conditions a model has to uncover rather than being handed upfront.

Two aspects of the framing stand out in the announcement. First, the tasks are described as long-horizon and as running at a minute-level timescale in an open world, which places them beyond short, tightly scoped problems. Second, the evaluation target is specifically multimodal models, and the hidden prerequisites are treated as part of what the benchmark measures.

Meituan's technical team says the evaluation shows that top multimodal models exhibit a capability gap on this class of tasks, one it describes as previously overlooked. The announcement does not name the models that took part, does not give the number of tasks or how they are scored, and provides no numeric results.

That is where the matter currently stands: MineExplorer has been introduced by its builders as an evaluation benchmark, together with the finding that leading multimodal models show a previously overlooked gap on such tasks. No further details about the benchmark itself, and no follow-up steps, have been disclosed in the material available so far.