China's AI model weekly call volume tops US for 22 straight weeks; anonymous 'Space Bunny' linked to MiniMax

September 28 — Shanghai-based AI giant MiniMax has released its latest large language model, the M3.1-Flash-Preview, now available for public beta testing, as Chinese AI models maintain their global lead in total token consumption for the 22nd consecutive week.

M3.1-Flash-Preview: Key Features

According to MiniMax, the new model supports:

• Native multimodal capabilities
• Million-token context windows
• Reliable performance for everyday development tasks, including bug fixes and full feature development

The model is designed to participate in the full development lifecycle — from issue identification and code implementation to testing and validation. It can handle edge-case logic, improve regression testing, and verify the impact of changes on existing functionality, forming a closed-loop development process from problem diagnosis to deliverable output.

The "Space Bunny" Connection

Following the release of M3.1-Flash-Preview, some developers have drawn comparisons to the recently popular anonymous model Space Bunny (also referred to as "Yutu" or "Jade Rabbit").

According to the latest rankings from global model aggregation platform OpenRouter, between September 21 and September 27:

• Total global AI model token usage reached 146 trillion tokens, up 13.18% week-over-week
• Chinese AI models accounted for 62.22 trillion tokens, down 7.77% from the previous week
• U.S. AI models accounted for 14.2 trillion tokens, down 0.07%

This marks the 22nd consecutive week that Chinese models have surpassed U.S. models in total weekly token usage.

Top 5 Global Models by Weekly Token Usage

| Rank | Model | Weekly Token Usage | Change | |------|-------|-------------------|--------| | 1 | DeepSeek V4.1 Flash | 19.6 trillion | +24% | | 2 | GLM 5.3 Flash (Zhipu) | 16.3 trillion | +16% | | 3 | Space Bunny (anonymous) | 13.9 trillion | — | | 4 | Hy4 preview (Tencent) | 9.64 trillion | -23% | | 5 | — | — | — |

Three of the top five models are Chinese. The anonymous Space Bunny model surged onto the榜单 during the Mid-Autumn Festival, ranking third.

Speculation on Space Bunny's Identity

Tokenizer tests have become a key clue in community discussions. Tests revealed that Space Bunny's token counting characteristics are consistent with MiniMax models. Some Reddit developers have further speculated that Space Bunny may be a preview version of MiniMax's M3.1 Flash. However, MiniMax has not officially confirmed this connection.

MiniMax H3: Open-Source Success

MiniMax recently open-sourced its next-generation multimodal generation model, MiniMax H3, which has achieved:

• Global No. 1 on the Artificial Analysis video editing benchmark
• Global No. 1 on the Arena text-to-video leaderboard
• Recognition from industry figures including Stable Diffusion founder Emad Mostaque and a16z partner Justine Moore
• Top popularity on Hugging Face, surpassing DeepSeek V4 Flash

CEO Commentary on Business Outlook

In a post-earnings conference call, MiniMax CEO Yan Junjie shared the following insights:

• Over the past two months, text model throughput per unit of compute has tripled
• The M3.1 model aims to reduce inference costs to roughly one-third of the M3's initial launch cost
• Cost savings will be split between passing savings to customers (expanding token volume) and improving gross margins

Yan noted that multimodal and voice models currently have relatively higher gross margins. As text model revenue grows rapidly and inference costs decline, he expects text business to become a key driver of overall margin improvement. He projected continued margin improvement in the second half of the year, with further upside in 2025.

He also predicted that as model capabilities and reliability improve, the business model may gradually shift from per-token billing to billing based on task outcomes and professional value.