Chinese artificial intelligence startup MiniMax is making a bold move in the global AI arms race. According to a Reuters report, the company is developing a large language model (LLM) with a staggering 2.7 trillion parameters. The model, which will be open-weight, could be released as early as the third quarter of this year.
What Is a Trillion-Parameter Model?
To understand the significance, it helps to know what a parameter is in AI. Think of parameters as the knobs and dials inside a neural network. The more parameters a model has, the more information it can store and the more complex tasks it can handle. A 2.7 trillion-parameter model is enormous—far larger than most publicly known models today.
For context, OpenAI's GPT-4 is believed to have around 1.7 trillion parameters, though the company has not confirmed the exact number. Meta's Llama 3.1, one of the most prominent open-weight models, has 405 billion parameters. MiniMax's model would dwarf both, making it one of the largest AI models ever built.
Open-weight means the model's trained parameters—the actual learned patterns—will be publicly released. This is different from open-source, which also includes the code used to train the model. Open-weight models allow developers and researchers to run the AI on their own hardware, but they cannot modify the underlying training process.
How MiniMax's Model Compares
If MiniMax delivers on its plans, its model would leapfrog China's current largest models. Reuters reports that Meituan's LongCat-2.0 and AI lab DeepSeek's V4-Pro both have 1.6 trillion parameters. MiniMax's 2.7 trillion would represent a roughly 70% increase in size.
This isn't just about bragging rights. Larger models can improve what AI researchers call reasoning—the ability to work through multi-step problems, understand context, and generate more coherent responses. They can also handle longer inputs, such as entire documents or extended conversations, without losing track of the topic.
But bigger models come with a catch: they are expensive to run. Training a model of this size requires vast amounts of computing power, typically from specialized AI chips. Running it for inference—actually using the model to answer questions or generate text—also demands significant hardware resources. That is where a technique called Mixture of Experts (MoE) comes in.
Mixture of Experts: Making Big Models Practical
Mixture of Experts is an architecture that splits the model into many smaller, specialized sub-models, or experts. When the model receives a query, it activates only a subset of these experts—the ones most relevant to the task. This keeps the computational cost much lower than if the entire model were used for every request.
MiniMax's use of MoE is not unusual; many modern large models, including Mixtral 8x7B and GPT-4, employ similar techniques. However, applying it at the 2.7 trillion-parameter scale is a significant engineering challenge. If successful, it could demonstrate that extremely large models can be both powerful and practical to deploy.
The broader AI industry is watching closely. The race to trillion-parameter models has intensified over the past year, with companies like Google, Meta, and various Chinese labs all pushing the boundaries. For investors, this signals that the demand for AI compute—and the chips that power it—is unlikely to slow down anytime soon.
What This Means for Investors
For everyday investors, the MiniMax news is a reminder that the AI sector remains highly competitive and capital-intensive. Companies that supply the hardware for training and running these models—such as chipmakers and cloud providers—stand to benefit from the ongoing arms race. Recent developments like AI chip startup SambaNova raising $1 billion and South Korean chip startup Rebellions planning an IPO underscore the market's appetite for AI infrastructure.
At the same time, the open-weight nature of MiniMax's model could accelerate the spread of powerful AI capabilities. This may benefit companies that build applications on top of such models, but it also raises questions about regulation and safety. Investors in AI-related stocks should keep an eye on how governments respond to the rapid advancement of open-weight models.
The broader market context also matters. Recent volatility, such as tech futures sliding on geopolitical tensions, shows that AI stocks are not immune to macro shocks. However, the long-term trend toward larger and more capable models suggests that the AI investment theme has staying power.
MiniMax has not confirmed a specific release date, and Reuters notes the model could arrive as early as Q3. Investors should watch for further announcements from the company and from other players in the trillion-parameter race. The outcome will shape not just the AI landscape, but also the fortunes of the companies and industries that depend on it.
