Chinese AI large model usage continues to hold the top spot globally, and Shanghai-based AI giant MiniMax has also released its latest large model product.

On September 28, reporters learned that MiniMax launched the text model M3.1-Flash-Preview and opened it for public testing. According to MiniMax, this model supports native multimodality and a million-token context window, providing stable and realistic productivity for daily development, and can reliably complete tasks such as bug fixing and full feature development.

The model can participate in problem localization, code implementation, and test verification, supplement processing logic for edge cases, improve regression testing, and verify the impact of changes on existing functions, forming a reliable development loop from problem identification to deliverable results.

Notably, after M3.1-Flash-Preview went live, some developers linked it to the recently popular anonymous model Space Bunny, also known as "Jade Rabbit."

Who is Space Bunny? According to the latest data ranking from global model aggregation platform OpenRouter, from September 21 to September 27, the total global AI large model usage was 146 trillion tokens, up 13.18% week over week.

Among the AI large models on the list, the weekly usage of Chinese AI large models reached 62.22 trillion tokens, down 7.77% week over week; during the same period, the weekly usage of U.S. AI large models was 14.2 trillion tokens, down 0.07% week over week.

This also marks the twenty-second consecutive week that Chinese large model weekly usage has exceeded that of the United States, ranking first globally.

Last week, among the top five in global usage, three were Chinese AI large models. Among them, DeepSeek V4.1 Flash ranked first for two consecutive weeks, with weekly usage reaching 19.6 trillion tokens, up 24% week over week; Zhipu GLM 5.3 Flash ranked second for two consecutive weeks, with weekly usage reaching 16.3 trillion tokens, up 16% week over week; Tencent Hy4 preview ranked fourth, with weekly usage reaching 9.64 trillion tokens, down 23% week over week.

The anonymous "Jade Rabbit" large model also surged onto the list during the Mid-Autumn Festival, ranking third, with weekly usage reaching 13.9 trillion tokens.

Around the true identity of Space Bunny, tokenizer testing has become a clue in community discussion and speculation. Related tests found that its token counting characteristics are consistent with MiniMax models.

Some Reddit developers further speculated that Space Bunny may be a preview version of MiniMax M3.1 Flash, though this claim has not yet been officially confirmed by MiniMax.

Previously, MiniMax officially open-sourced its new-generation general multimodal generation model MiniMax H3, which ranked first globally on the Artificial Analysis video editing evaluation list and the Arena image-to-video leaderboard, and received praise from well-known industry figures such as Stable Diffusion founder Emad Mostaque and a16z partner Justine Moore.

On the open-source community Hugging Face, MiniMax H3 once rose to first place in popularity, surpassing DeepSeek V4 Flash.

Earlier, during a conference call after its earnings report, MiniMax CEO Yan Junjie said that over the past two-plus months, the per-unit compute throughput of the text model increased threefold, and the goal of M3.1 is to reduce inference costs to about one-third of what they were when M3 first launched; after costs decline, part of the benefit will be passed on to customers to expand token scale, and part will be retained to improve gross margin.

Yan Junjie said that currently the gross margins of multimodality and voice are relatively higher, while the revenue share of the text model is rising rapidly and inference costs are falling quickly. He expects the text business to become an important driver of overall gross margin improvement, with gross margin continuing to improve in the second half of the year and still having room for further improvement next year.

He also judged that as model capabilities and reliability improve, the business model may gradually shift from billing by token to billing by task results and professional value.