MiniMax has released its first semi-annual report since its Hong Kong listing, with revenue increasing 283.1% year-over-year to 1.5 times its full-year 2025 target, and gross profit rising 464.8%. In August, annualized recurring revenue (ARR) surpassed $800 million, with B2B revenue accounting for 80%. In July, token consumption reached 20 times the level of January. Founder Yan Junjie has proposed a technical roadmap of “minimizing inference costs while maximizing intelligence,” achieving a reduction in token computation cost for million-word-long texts to approximately 1/20 of traditional full-attention mechanisms through its proprietary MSA architecture. This enables an optimal balance between intelligence and cost under limited computing resources, offering a new pathway for Chinese large models to compete globally.

Article author and source: Zhixidong

On August 26, China's leading AI large model company, MiniMax (Xiyu Technology), released its first half-year report since its Hong Kong stock listing.

The financial report shows that MiniMax's revenue increased by 283.1% year-over-year, with its half-year revenue reaching 1.5 times the full-year revenue of 2025; gross profit increased by 464.8% year-over-year.

At the earnings call, Yan Junjie, founder, chairman of the board, CEO, and CTO of MiniMax, revealed that in August, MiniMax’s ARR surpassed $800 million, with B2B revenue accounting for 80%, and token consumption in July reached 20 times the level of January.

Beyond the data, what caught my attention was a statement Yan Junjie made during last night’s earnings call: “Only by minimizing unit intelligent cost can we maximize intelligent capability.”

Over the past two years, large models have essentially followed two paths: either scaling up parameters according to the scaling law to boost intelligence—at the cost of high expenses—or reducing efficiency to make models more accessible, but often falling short in terms of intelligence.

The industry seems to treat this as a case of mutually exclusive choices.

Why did MiniMax choose a path that goes against industry intuition? Can this path actually succeed?

Intelligence and cost—why can’t both be achieved?

The Scaling Law has driven the narrative around AI over the past few years: larger parameters equal greater intelligence.

Following the scaling law, model parameters have now reached the 2-3 trillion range, and model intelligence has risen accordingly, leading many to exclaim that AGI is just around the corner.

Behind the prosperity, this rule has also planted a bomb: when models reach the trillion-parameter scale, inference costs begin to become significant.

NVIDIA CEO Jensen Huang stated at the 2026 GTC conference: "The era in which training costs were the primary driver has ended. Inference is now the dominant workload, and it is expanding rapidly."

The inflation of inference costs, combined with the rise of agent workloads, is driving AI from one-off Q&A to multi-step autonomous execution. A single user request may now trigger dozens or even hundreds of model calls, causing the industry’s compute bills to rise at an exponential rate.

As a result, the industry gradually split into two camps:

On one side are small models with fast speed and low cost, but average intelligent performance.

Google is a classic example of this path. Between July and August 2026, Google launched four lightweight models in quick succession: Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber.

3.5 Flash-Lite is the fastest and most cost-effective model in the 3.5 series, designed for high-throughput workloads and smaller tasks within large AI agent systems. The output token usage of Gemini 3.6 Flash has decreased by 17% compared to its predecessor.

But the cost is evident: Google has still not launched its flagship model, Gemini 3.5 Pro, which was originally scheduled to be unveiled at the May I/O conference.

On the other side are large models with high intelligence but expensive deployment and rising costs.

Domestic large model manufacturers are racing down this path. In July 2026, Moonshot AI released Kimi K3, with a parameter scale of approximately 2.8 trillion, making it the largest open-source model in the world by parameter count.

Alibaba subsequently launched Qwen3.8 Max with approximately 2.4 trillion parameters; Baidu's Wenxin 5.0 also reached 2.4 trillion parameters; DeepSeek V4-Pro has 1.6 trillion parameters.

The performance gains from increased parameter scale are visibly evident. Kimi K3 outperforms Claude Fable 5 and GPT-5.6 Sol on multiple Agent leaderboards. Qwen3.8 Max is described as one of the largest and most powerful open-source models by parameter scale.

But costs have also risen accordingly.

The API pricing for Kimi K3 has increased: non-cached input costs have risen from RMB 6.5 per million tokens to RMB 20, a 208% increase, while output costs have increased from RMB 27 to RMB 100, a 270% increase. The official recommendation is to deploy using super-nodes composed of 64 or more accelerator cards.

Just two days after launch, Moon Shadow announced it would suspend new consumer subscriptions and allocate all computing power to serve existing users.

The more advanced the intelligence, the larger the model parameters, the higher the cost of invocation, the greater the barrier to deployment, and the more scarce the computing power becomes... it seems like an unsolvable deadlock.

MiniMax, no compromises

For the dilemma of choosing between cost and intelligence, MiniMax’s choice is: I want both.

Yan Junjie clearly laid out MiniMax’s entire logic during the earnings call:

Computing power is limited for every company. We aim to achieve “Intelligence with Everyone” by “Minimizing Inference Costs and Maximizing Intelligence.” This has been our consistent technological path and our original entrepreneurial vision.

“Minimize the Inference Cost, Maximize the Intelligence” means “Minimize inference cost, maximize intelligence.” MiniMax places cost reduction and intelligence enhancement within the same narrative framework.

Why can these two things be mentioned together? Because MiniMax has a different understanding of "reasoning":

First, reasoning is the means of production for building the next generation of intelligence.

Previously, reasoning was considered “using a model,” while training was seen as “building a model.” But now, core processes such as synthetic data generation, reinforcement learning rollouts, model evaluation and validation, and agent-environment interactions all fundamentally run inference workloads.

The post-training phase is accounting for an increasingly larger share of total training compute. Inference is no longer a workload that only emerges after training concludes—it is a continuous consumption throughout the entire training process.

Therefore, inference efficiency determines not only the pricing potential of APIs but also the extent to which the next generation of models can be trained.

Second, optimizing inference costs is a necessary condition for intelligent iteration.

Computational power has a physical limit, while model sizes continue to grow. If inference costs cannot be reduced, they will increasingly strangle further advances in intelligence.

Within a limited computational budget, how effectively each dollar is converted into meaningful intelligent output determines how far model iteration can go.

Third, pursuing the highest level of intelligence at the lowest possible unit cost are two sides of the same goal.

The industry has traditionally viewed "intelligence-first" and "cost-first" as two separate paths. But without bringing down inference costs, no matter how advanced the intelligence, it cannot be practically implemented.

Simply put, MiniMax treats cost optimization and intelligent enhancement as one and the same. From the very first day of model design, inference efficiency is treated as a core metric, integrated throughout the entire development lifecycle.

As Yan Junjie put it: “MiniMax does not seek to trade off intelligence for cost. Achieving higher levels of intelligence is the goal; improving cost-effectiveness through continuous optimization of inference cost and efficiency is the means.”

MiniMax has always adhered to this original mission, building a solid technological foundation.

The most critical aspect is its proprietary MSA (Sparse Attention) architecture. In simple terms, it reduces the computational cost per token for million-token-long texts to approximately 1/20 of that of traditional full attention mechanisms, meaning M3’s per-token computation cost under a 1M context is only 1/20 of its predecessor’s.

The key to this breakthrough is that the total number of parameters determines the model’s upper limit of knowledge capacity, while the number of activated parameters determines the inference cost per token. These two factors can be decoupled: increasing parameter scale does not necessarily lead to a proportional increase in inference cost.

Therefore, the M3 can enhance intelligence without increasing price, while the M3 Pro can achieve a parameter scale of approximately 3 trillion while advancing reinforcement learning and long-horizon task training.

At the infrastructure level, MiniMax achieves an ETTR (Effective Training Time Ratio) of 97%, making it one of the first independent large model companies in China to establish a large-scale, long-term stable computing power system.

We have built a computing power supply network through a combination of self-managed clusters, cloud service providers, and TokenFactory. The current and planned computing power is sufficient to support the continuous iteration of text and video models at the 3T scale.

Conclusion: Without a computational advantage, how can domestic models catch up?

Whenever the gap between Chinese and U.S. large models is discussed, hardware is always an unavoidable topic.

In the past, it was commonly assumed that whoever had more advanced chip manufacturing processes and higher computational power could build better models. By this logic, Chinese AI would always remain several months to half a year behind U.S. AI.

Does it really have to be this way?

MiniMax has charted a new path: Minimize the Inference Cost, Maximize the Intelligence.

Reducing the cost per unit of intelligence also raises the upper limit of intelligence. Because today’s post-training, synthetic data, and reinforcement learning fundamentally rely on inference workloads, higher inference efficiency enables more experiments to be conducted with the same computational power, thereby raising the ceiling of intelligence.

Whoever can produce higher intelligence at a lower cost can push the boundaries further within limited computing power, enabling higher intelligence to enter a broader range of everyday life.

This may be the most promising path for domestic large models to compete globally—and MiniMax has already begun to validate it.