NextFin News — MiniMax Group’s first-half results show a model company beginning to convert technical progress into commercial traction. Revenue reached about $117 million in the six months ended June 30, more than triple the year-earlier figure and already higher than the company’s entire 2025 total.

The more telling shift sits in the mix. Enterprise and open-platform services now account for most of the top line, overseas markets supply the majority of sales, and gross profit has expanded sharply even as the company continues to post sizable losses.

Open-platform and other enterprise AI services generated $73.9 million, up more than sevenfold, and contributed 63 percent of revenue, compared with 30 percent a year earlier. Revenue from AI-native products doubled to $42.6 million. Overseas markets accounted for 61 percent of sales. Research and development spending rose 139 percent to $297 million—still several times revenue—while sales and distribution costs fell nearly 18 percent as adoption relied more on organic demand than paid promotion. The net loss narrowed 11 percent to $358 million. Adjusted net loss stood at $293 million, and the adjusted loss rate declined from roughly 456 percent to about 251 percent.

Founder and chief executive Yan Junjie described the underlying constraint in plain terms. Intelligence can keep improving, he said, but energy and compute are finite. Token consumption in July reached twenty times the January level. The company’s priority remains lowering the cost of delivering higher intelligence to more users rather than maximizing model size for its own sake.

Product updates during and after the period reflect the same emphasis. The M3 model retained earlier pricing while improving capability and drawing additional enterprise and developer workloads. Usage is moving from single-turn human queries toward multi-step agent workflows that generate chains of model calls, tool use and sub-tasks. That pattern is pushing token volume ahead of growth in users or messages. After the reporting period the company released an open-source video model, H3, that quickly attracted derivative models and high download volumes.

MiniMax said its models and products now serve more than two million enterprise customers and developers—roughly ten times the figure at the end of 2025—and more than 300 million individual users across more than 230 countries and regions. Shares closed little changed on the results day at about HK$303, implying a market value near HK$106 billion.

From capability contests to serving economics

The figures sit inside a broader transition across the model industry. For several years developers competed primarily on benchmark scores, parameter counts and headline capabilities. That phase produced rapid technical gains and heavy capital consumption. As models have become good enough for many practical tasks, attention has shifted to cost per token, utilization of serving infrastructure, the durability of enterprise workflows, and the ability to monetize usage driven increasingly by agents rather than individual chat sessions.

Enterprise and API revenue has emerged as the clearest commercial signal. When businesses embed models in customer service, content systems, internal tools or multi-agent processes, consumption becomes recurring and measurable. Consumer applications can still build brand and data advantages, yet they have proven harder to convert into high-margin, predictable revenue at scale. MiniMax’s move toward a majority enterprise mix mirrors a pattern visible among several model companies that publish segmented results: the fastest growth and the most defensible pricing power appear where models function as infrastructure rather than as destinations.

Overseas revenue shares above 50 percent are becoming more common among Chinese model providers that offer competitive open weights or aggressive API pricing. Global developers and startups remain price-sensitive and willing to switch among suppliers that deliver acceptable quality at lower cost. That dynamic rewards continuous efficiency gains in training and inference. It also exposes strategies that depend too heavily on a single domestic market.

Losses remain structural for most pure model companies. Training, talent and serving capacity still require large outlays relative to current revenue. The more useful signal in recent results is the relationship between revenue growth and cost growth. When sales expand faster than research spending and commercial costs are contained, the path to narrower losses becomes visible even if absolute break-even remains distant. Gross-margin expansion is an early indication that utilization and pricing are beginning to improve.

The pressures that will define the next phase

Three forces will shape the period ahead. Agent-driven workloads raise token intensity per task; providers that cannot reduce unit costs will see margins compressed even as volume grows. Open-source releases accelerate adoption and ecosystem effects but intensify price competition and reduce differentiation on raw capability alone. Enterprise buyers are growing more sophisticated about total cost of ownership, latency, reliability and data handling, moving procurement conversations away from pure benchmark comparisons.

MiniMax’s first-half results show one company navigating that shift: rapid revenue growth, a decisive turn toward enterprise and API income, rising overseas contribution, improving gross profit, and continued heavy investment tempered by early operating leverage. Similar patterns are appearing, to varying degrees, among model companies that have chosen public markets or detailed disclosure. The industry is no longer judged solely by who trains the most impressive model. It is increasingly judged by who can deliver usable intelligence at a cost and scale that enterprises will pay for repeatedly. That test is only beginning.