MiniMax developing world's largest open LLM with 2.7 trillion parameters

Chinese AI startup MiniMax is developing a large language model with 2.7 trillion parameters, reports Reuters citing an informed source. This is the largest AI model with open weights released by a Chinese company, and possibly the largest open model in the world.

How large the MiniMax model really is

2.7 trillion parameters — a figure that places the future MiniMax model in the top echelon of known language models by declared size. For comparison: Chinese DeepSeek, whose DeepSeek V3 model (671 billion parameters) once made a splash on the market of open AI models, works with a model nearly four times smaller in number of parameters than the one being developed by MiniMax.

• Developer — Chinese AI startup MiniMax
• Declared model size — 2.7 trillion parameters
• Source of information — Reuters citing an informed source
• The model is claimed to be the largest open AI model released by a Chinese company

Who is MiniMax and why you should know about it

MiniMax is one of the so-called Chinese "six tigers" of AI startups founded in Shanghai and releasing their own large language, voice, and video models under the brands abab and MiniMax-01/M-series. The company has already published open models and API services for text, speech, and video generation and competes within China with DeepSeek, Alibaba (Qwen model line), Moonshot AI (Kimi), and Zhipu AI. The development of a model of such scale as 2.7 trillion parameters requires significant computing power and speaks to the serious ambitions of the company to establish itself in the top echelon of global AI laboratories, not just the internal Chinese market.

Why the race for model size has intensified again

Large language models of such scale are almost always built on the MoE (mixture-of-experts) architecture, where only a portion of the total number of parameters is involved in processing each specific query — this allows scaling up the total model size without a proportional increase in computational costs for inference. The race for the number of parameters among Chinese AI laboratories has become one of the main fronts of competition with American laboratories in the market of open weights: after DeepSeek's open models made a serious impression on the industry with their quality-to-training-cost ratio, interest in Chinese open models soared among developers around the world seeking an alternative to closed APIs from OpenAI, Anthropic, and Google.

Open weights mean that after publication, the model can be downloaded, fine-tuned for your own tasks, and run on your own infrastructure without dependence on a specific provider's API — this is how Chinese laboratories fundamentally differ from most leading American developers, who keep the weights of flagship models closed.

What this means

If Reuters' data is confirmed, MiniMax will cement its status as the developer of the largest open-source language model in the world — another signal that it is Chinese AI laboratories, not American ones, that are currently setting the pace in the race for the scale of models with open weights.