MiniMax has taken the lead in the self-evolution of domestic large models.

Nowadays, the hot topic isn't about new models topping the charts or securing multi-million-dollar industry contracts. Instead, it boils down to two words: price increases.

Just yesterday, Alibaba Cloud and Baidu Intelligent Cloud announced simultaneously that starting April 18, there would be price adjustments for their AI computing power and storage products, with maximum increases reaching 34%. Prior to this, Amazon AWS, Microsoft Azure, and Google Cloud had already raised prices for their AI-related services, with some specific projects seeing increases of up to 100%. Cloud providers worldwide have collectively entered a cycle of computing power price hikes, leaving none unaffected.

In other words, developing large models now is akin to driving on a highway with no exits. You dare not ease off the accelerator, lest you be instantly overtaken by the cars behind you. However, fuel prices keep rising, and the financial resources in your tank are dwindling faster and faster. No one knows if they can make it to the next service station.

This is no exaggeration. If you take a look at Alibaba's Q3 2026 financial forecast and Tencent's recently released financial report, you'll notice their biggest commonality: significantly increased and sustained investment in AI infrastructure and large model R&D. Of course, this has inevitably slowed down revenue growth for these former internet giants.

Almost simultaneously, Shanghai-based large model startup MiniMax unveiled its new flagship large model, M2.7. Unlike the industry's usual focus on "record-breaking parameter scales" or "comprehensive leadership on authoritative benchmarks," the core label of this model is "self-evolution."

According to the official introduction, this is the world's first commercially viable self-evolving large model. It can deeply participate in the entire process of its own training and optimization, handling 30% to 50% of the workload in certain R&D scenarios.

On one hand, the entire industry is anxious about computing power costs and R&D efficiency, sinking deeper into an arms race. On the other hand, a startup founded just over four years ago has transformed large models from mere R&D tools into active participants—even the main drivers—of the R&D process. This move has hit the industry's weak spot.

01

Three Years of Competition: The Hidden Challenges Behind Large Models

The large model industry has now undergone three years of development, with nearly all players trapped in the same arms race.

When ChatGPT ignited the generative AI wave in 2023, the industry's competitive focus was on parameter scale. From hundreds of billions to trillions of parameters, the expansion rate of large models far outpaced Moore's Law, as if parameter scale were the sole metric for measuring a large model's capability. Whoever had more parameters stood at the top of the industry.

Soon, the parameter race reached its limits. It became clear that increasing parameter scale yielded diminishing returns in capability growth while exponentially driving up computing power demands. The industry's competitive focus then shifted to securing computing power resources. The supply-demand gap for high-end AI chips widened, with domestic large model companies locking in long-term computing power contracts. The industry even faced a "chip shortage" crisis, with some startups renting entire data center cabinets to ensure stable computing power.

By 2025, the pure computing power arms race had also stalled, and the industry's battleground shifted to deployment scenarios. Tech giants began integrating large models with their cloud services, hardware products, and ecosystems, vying for first-mover advantages across industries. Competition expanded from technological R&D to comprehensive contests in ecosystems, channels, and customer resources.

But this three-year competition has come at an increasingly heavy cost, beginning to hinder the industry's development.

The pressure is even more visible for startups. Many large model entrepreneurs I've spoken with share similar experiences: half of their initial funding goes to cloud providers to secure computing power, most of the remaining funds pay algorithm teams' salaries, leaving little for product refinement and scenario deployment. The industry resembles a high-speed treadmill—everyone must keep running, or they'll be overtaken. But the faster they run, the higher the costs, creating a vicious cycle of "more competition, higher costs; higher costs, more competition."

The root of this problem lies in the fundamental bottleneck of large model R&D efficiency.

Traditional large model R&D follows a standardized, human-driven process. From initial data cleaning and labeling to model architecture design, pre-training parameter adjustments, and subsequent fine-tuning, alignment, evaluation, and bug fixes, every stage requires heavy involvement from algorithm engineers, data labelers, and product managers.

A large model with hundreds of billions of parameters often requires a team of several hundred people and three to six months to go from project initiation to official release, consuming hundreds of millions in computing power costs. Even the fastest-iterating top players need at least two months for a major version update.

More frustratingly, this "artisanal" R&D model is experiencing clear diminishing marginal returns. From GPT-3 to GPT-4 to GPT-5, OpenAI has invested more computing power and manpower with each iteration, yet ordinary users perceive increasingly limited capability improvements.

The same goes for domestic players. From 2023 to 2026, hundreds of large models have been released in China, with growing parameter scales and training data volumes. However, truly transformative capabilities that reshape industry processes or deliver disruptive experiences remain rare. Many model iterations merely improve benchmark scores, rarely translating into user-perceptible upgrades.

02

Self-Evolution: A New Exit for the Large Model Industry

From this perspective, MiniMax's M2.7 points to a brand-new path.

M2.7's core breakthrough isn't about increasing parameter scale or optimizing specific scenario capabilities—it's about reconstructing the large model R&D paradigm.

Before M2.7, large models could only serve as auxiliary tools in their own R&D processes. Algorithm engineers might use them to write training-related code or perform simple data cleaning, but core tasks like model architecture design, training process control, and alignment optimization still required human intervention. Large models remained objects of R&D, not participants. The entire R&D process was human-driven.

M2.7's "self-evolution" capability, however, enables large models to deeply participate in their own R&D lifecycle for the first time.

According to MiniMax's official disclosure, M2.7 can handle tasks such as data screening and cleaning, training data construction, model architecture iteration and optimization, in-training parameter adjustments, alignment and evaluation, and even optimizing its own inference code. In certain R&D processes, M2.7 can handle 30% to 50% of the workload, with R&D personnel only needing to set top-level goals, review key stages, and validate final results.

This change essentially shifts the large model R&D model from "human-driven model iteration" to "model-driven model iteration."

The most immediate impact is improved R&D efficiency and reduced costs.

Whereas a major model iteration previously required a 200-person algorithm team and three months, it can now be completed with fewer personnel and a shorter cycle, significantly cutting labor and time costs. Computing power costs also decrease, as M2.7 autonomously optimizes training processes and parameters, improving computing power utilization efficiency and reducing resource consumption for the same training tasks. For cash-strapped startups, this efficiency boost directly expands their survival space.

Many may wonder: does involving large models in their own R&D compromise foundational capabilities?

According to official information, it does not. M2.7's programming capabilities now match OpenAI's GPT-5.3-Codex, while its core strengths in multimodal understanding, long-context processing, and logical reasoning place it among China's top tier. Self-evolution hasn't come at the expense of foundational capabilities—instead, autonomous model optimization has enhanced them.

For a startup founded just over four years ago, such achievements are no accident. Founded in 2022, MiniMax was among China's first large model startups, having released multiple general-purpose large model versions. It has accumulated mature technical capabilities in multimodal generation and long-dialogue scenarios, along with extensive model R&D data and experience—all of which laid a solid foundation for self-evolving large model development. With this release, MiniMax has leaped from China's second tier of large model companies into the first tier of technological innovation.

The reason M2.7 has caused such an industry stir is that it breaks the competitive logic of the large model industry over the past three years.

For the past three years, large model competition has essentially been a resource contest. Whoever could secure more computing power, assemble larger algorithm teams, or invest more capital gained a competitive edge. Tech giants leveraged their financial, computing power, and ecological advantages to dominate this competition, leaving startups to seek niches in specific scenarios while struggling to compete head-on with giants on core general-purpose model technologies.

The industry's Matthew effect has become increasingly pronounced, with top players monopolizing most computing power, talent, and market share, while smaller players face shrinking survival space.

But self-evolving large models change this game. When models can develop themselves, resource importance declines relative to technological innovation. Startups no longer need to compete with giants on computing power, capital, or team size—they just need core technological breakthroughs to achieve faster iteration with fewer resources, securing their place in the competition.

Similar to DeepSeek's impact in its time, this opens a new track for an industry trapped in internal competition, making technological innovation the core of industry competition once again.

03

The New Chapter of the Large Model Industry Has Just Begun After the Gunshot

The same logic is unfolding globally.

According to the latest research analysis by Tianfeng International Securities analyst Ming-Chi Kuo, NVIDIA is increasing its investment in AI inference architecture company Groq. Groq's LPU shipment plans have seen significant upward revisions, with cumulative shipments expected to reach 4-5 million units from 2026-2027—over 10 times previous estimates. Meanwhile, NVIDIA plans to increase LPU configuration per AI cabinet from 64 to 256 units, expanding storage capacity while maintaining ultra-low-latency inference performance to meet explosive AI inference demand.

NVIDIA's strategy aims to address the AI industry's new bottleneck: as large model deployment scales up, inference computing power consumption is growing exponentially—one of the core reasons behind global cloud providers' price hikes.

In other words, overseas giants are upgrading specialized chips to improve inference efficiency and reduce costs from the hardware side. MiniMax's self-evolving model, however, approaches the problem from the software and model side, using autonomous model optimization to enhance inference efficiency and reduce hardware dependency.

Industry sources claim that OpenAI's next-generation large model will focus on self-optimization and iteration, attempting to involve large models deeply in their own R&D processes. However, OpenAI has yet to release a commercially viable self-evolving large model. MiniMax's M2.7 is thus the world's first commercially available self-evolving large model, meaning China has taken the lead globally in large model R&D paradigm innovation.

Beyond R&D efficiency and competitive logic, self-evolving large models may also reshape the entire AI industry's supply chain.

Sources reveal that OpenAI and AWS simultaneously disclosed new collaboration progress, providing exclusive AI services to U.S. government agencies within regulatory frameworks to further expand their government-enterprise market share. Meanwhile, Microsoft's cloud service cooperation disputes with OpenAI and Amazon continue to escalate, with rumors suggesting Microsoft may take OpenAI and Amazon to court over exclusive AI cloud service agreements.

Behind these disputes lies the increasingly tense game between large model companies and cloud providers. Historically, large model companies have heavily relied on cloud providers' computing power, giving cloud providers core bargaining power over the supply chain and even allowing them to influence large model companies' development pace through computing power supply. The domestic large model industry follows a similar pattern, with most companies dependent on computing power from Alibaba Cloud, Baidu Intelligent Cloud, Huawei Cloud, and other leading providers, often finding themselves in passive positions during cooperation.

However, as self-evolving large models deploy, large model companies' dependency on computing power will significantly decrease. This means they'll gain more bargaining power in interactions with cloud providers, potentially altering the supply chain's profit distribution. For cloud providers, mere computing power sales can no longer sustain past high growth—they must transition toward upper-layer ecological services and solutions, driving transformation and upgrading across the cloud services industry.

Three years have passed since ChatGPT ignited the generative AI wave. During this time, the large model industry has evolved from wild growth to an arms race, and now to deployment competition, reaching a critical crossroads.

The escalating costs of computing power, coupled with diminishing marginal returns, have prompted the entire industry to recognize that development models heavily reliant on resources have reached their limits. The large-scale model industry now stands in need of a fresh technological revolution to overcome the existing bottlenecks. MiniMax's self-evolving large-scale model, M2.7, has fired the opening salvo in this transformative endeavor.

This article is an original work contributed by Xinmou.

— END —