The announced figure

According to IT之家, 面壁智能 (ModelBest) announced that, in November 2026, the MiniCPM model family surpassed 70 million downloads globally. This is a company statement, reported by the outlet, not independently verified data: the material does not specify on which platforms or repositories the count is calculated, nor how double-counting of downloads is avoided.

The number should be read alongside ModelBest’s origins: founded in Beijing in August 2022, the company emerged from an incubation of Tsinghua University’s natural language processing laboratory. CEO Li Dahai was a partner and CTO of Zhihu; chief scientist Liu Zhiyuan is a professor in Tsinghua’s computer science department; CTO Zeng Guoyang, at 22, led the training of CPM-1 as chief engineer, described as the first Chinese large language model.

From version 4.7 to Galaxy integration

The download announcement comes about a month after the October launch of MiniCPM-V 4.7, a new generation of on-device multimodal model offered in two sizes, 35B-A3B and 1B. The 35B-A3B variant is presented as the first MoE model in the MiniCPM-V series and, according to the official description cited by the source, is the first open-source MoE capable of running directly on-device without loss of quality. The reference provided is a decoding speed above 13 tokens per second on iPhone 17 Pro Max with only 3 GB of memory, thanks to LES acceleration technology.

On the hardware side, in July 2026 Samsung presented the Galaxy Z Fold8, Z Fold8 Ultra and Z Flip8 foldables; on that occasion ModelBest stated that MiniCPM models power Galaxy AI. The source presents the move as the first entry of a Chinese on-device model into the global flagship line of a major international manufacturer. This, too, is a company statement: the actual scope of the integration, the functions involved and the exact models used are not specified.)

The combination of the download count and presence on a flagship line indicates a strategy oriented toward distribution on end devices rather than only cloud access. It is a positioning consistent with the two sizes of MiniCPM-V 4.7: a 1B version designed for memory constraints, and a 35B-A3B version for devices with greater resources. The material does not, however, allow us to say which of the two is used on Galaxy AI.

What remains to be verified

Elements are missing to assess the robustness of the claims: no download-counting methodology is indicated, nor a breakdown by region or channel. It is unclear whether the 70 million include weights downloaded from public repositories, integrations in third-party applications, or distributions through partnerships. Even the claimed performance record for the 35B-A3B MoE on iPhone is based on a measure provided by the company, without independent benchmarks in the material.

The most solid element is chronological: the announcement comes after the launch of MiniCPM-V 4.7 and after the July statement on integration with Samsung. Otherwise, the news should be treated as a corporate communication, useful for outlining ModelBest’s trajectory but not sufficient to confirm market share or the technical quality of the models.

Source and attribution: this original summary was prepared by AICronaca based on the report by IT之家 ↗. The statements are attributed to the source; any information limitations: All data on downloads, performance and integration come from ModelBest statements reported by IT之家; no independent sources, counting methodology or confirmation from Samsung are available..