China’s MiniMax Unveils AI Reasoning Model That Outperforms DeepSeek
China’s artificial intelligence (AI) industry has experienced an extraordinary surge in innovation and competitiveness over the past five years. Once seen as trailing behind American firms in developing cutting-edge AI models, Chinese companies have rapidly closed the gap, with several now challenging global leaders in specialized domains such as multimodal processing, large language models (LLMs), and increasingly, reasoning AI. In this context, reasoning models—those designed to emulate human-like logical thinking, step-by-step problem-solving, and contextual understanding—have become a critical battleground. As enterprises, governments, and research institutions demand AI systems capable of more advanced cognitive functions, the ability to develop superior reasoning models is emerging as a key differentiator.
Against this backdrop, Shanghai-based MiniMax has announced a significant milestone: the release of a new AI reasoning model that it claims outperforms DeepSeek, one of China’s leading reasoning-capable LLMs. This announcement has generated considerable attention within both domestic AI circles and the international AI research community. DeepSeek, backed by a powerful coalition of Chinese tech investors and talent from top academic institutions, had previously set the benchmark for AI reasoning performance within China. MiniMax’s new model not only challenges this status quo but, according to internal testing and preliminary benchmarks, surpasses DeepSeek in a variety of reasoning-heavy tasks.
This development underscores several broader trends that are reshaping the global AI landscape. First, it highlights the growing sophistication of China’s AI companies, which are increasingly capable of delivering technical breakthroughs rather than merely replicating Western models. Second, it points to the strategic importance Beijing places on mastering reasoning AI as a pillar of national competitiveness in sectors ranging from manufacturing to healthcare and military technology. Third, the MiniMax-DeepSeek rivalry illustrates the dynamic and often volatile nature of China’s AI startup ecosystem, where fast-moving, well-funded companies are constantly seeking to leapfrog one another.
This blog post provides an in-depth examination of MiniMax’s new reasoning model and what it means for the broader AI field. We will begin with an overview of MiniMax’s rise within China’s AI ecosystem, then delve into the technical innovations underlying the new model. We will compare its performance against DeepSeek and other domestic models, assess the strategic implications for China’s AI ambitions, and conclude with a forward-looking analysis of where this competition may head next. Through this exploration, we aim to provide readers with a comprehensive understanding of a pivotal moment in the global evolution of reasoning AI.
The Rise of MiniMax in China’s AI Ecosystem
Founded in 2021 by a group of seasoned AI researchers and former executives from leading Chinese technology firms, MiniMax has quickly ascended to prominence within China’s increasingly crowded AI landscape. In a market dominated by heavyweights such as Baidu, Tencent, Alibaba’s DAMO Academy, and SenseTime, MiniMax has carved out a distinctive niche through a relentless focus on reasoning AI and cognitive capabilities that go beyond standard natural language processing. The company’s meteoric rise is emblematic of the broader dynamism shaping China’s AI ecosystem—a highly competitive, state-supported environment where startups with strong technical foundations and clear strategic vision can rapidly achieve national relevance.
At its inception, MiniMax positioned itself as an independent AI innovator with a strong emphasis on R&D. The company attracted early venture funding from several prominent investors, including Sequoia China and Prosperity7 Ventures. These funds were strategically allocated toward building an elite research team composed of AI scientists from Tsinghua University, Peking University, and Chinese Academy of Sciences, alongside engineers with industry experience from Baidu’s ERNIE team and Alibaba Cloud. The combination of academic rigor and industry pragmatism has allowed MiniMax to iterate rapidly and push the frontiers of reasoning AI, an area often overlooked in China’s earlier wave of LLM development.
By 2023, MiniMax had released its first generation of LLMs capable of multi-step reasoning and complex query resolution, earning early recognition from both enterprise clients and academic reviewers. Rather than pursuing broad general-purpose capabilities in the style of ChatGPT or Claude, MiniMax chose to specialize in verticals where reasoning is mission-critical: financial modeling, legal document analysis, scientific research support, and industrial process optimization. This specialization resonated with key sectors of China’s economy, particularly in the context of Beijing’s broader strategic goal of achieving AI self-sufficiency in sensitive domains.
An important factor in MiniMax’s growth has been its alignment with China’s national AI strategy. The Ministry of Science and Technology, along with provincial governments in Shanghai and Shenzhen, has identified reasoning AI as a “strategic emerging technology” with implications for national security, advanced manufacturing, and healthcare innovation. MiniMax has successfully secured government grants and participated in several state-sponsored pilot programs aimed at integrating reasoning AI into public sector applications. This relationship with policy-making bodies has provided both financial resources and invaluable access to high-quality proprietary data, further enhancing the model’s capabilities.
MiniMax’s commercial strategy has also been notable for its agility. Rather than building a purely consumer-facing AI assistant, the company has focused on licensing its reasoning models to enterprise partners across multiple industries. These partners range from major Chinese banks seeking AI-driven fraud detection and risk analysis to manufacturers implementing AI for predictive maintenance and supply chain optimization. Through this enterprise-first approach, MiniMax has cultivated a client base willing to pay premium licensing fees, giving the company a more sustainable revenue model compared to some of its LLM competitors.
In parallel, MiniMax has built a strong collaborative network with universities and research institutes, contributing to open research papers and participating in shared benchmark evaluations. The company’s transparency around model performance and its willingness to subject its models to rigorous academic scrutiny have earned it credibility in China’s AI research community. This stands in contrast to more closed corporate labs, which has helped MiniMax attract top AI talent and foster an innovative internal culture.
As of 2025, MiniMax ranks among the top five Chinese AI firms in terms of model capability and enterprise adoption, positioning itself as a credible rival to DeepSeek, Baidu ERNIE, and SenseTime’s LLaMA-based models. Its new reasoning model, which claims to outperform DeepSeek on several critical benchmarks, represents both a technical achievement and a strategic gambit—one that could further elevate MiniMax’s standing in the global AI race.
The Strategic Importance of Reasoning AI in China’s National Plans
As China intensifies its efforts to become a global leader in artificial intelligence, reasoning AI has been identified as a critical focus area within the country’s national AI development strategy. Unlike early large language models that were primarily designed for text generation and conversational fluency, reasoning AI models—capable of multi-step logic, causal inference, and decision support—are now seen as vital enablers for the next wave of AI-driven economic and industrial transformation.
The 14th Five-Year Plan explicitly names “advanced cognitive AI” and “explainable AI” as priority research areas. Policy documents issued by the Ministry of Industry and Information Technology (MIIT) and the National Development and Reform Commission (NDRC) further underscore that reasoning capabilities are necessary for applications in smart manufacturing, legal AI, healthcare diagnostics, and AI-assisted scientific research. The aim is not merely to automate repetitive tasks but to create AI systems that can serve as trusted advisors and decision-support agents across strategic sectors of the economy.
In this context, MiniMax’s MM-ReasONE aligns perfectly with national priorities. Its superior reasoning and interpretive skills position it as an ideal candidate for deployment in government-backed pilot programs related to AI-powered public administration, judicial assistance, and compliance monitoring. Moreover, MM-ReasONE’s performance in regulated industries such as finance and healthcare enhances China’s ability to develop AI solutions that can operate within the country’s complex legal and policy frameworks—an area where many foreign models fall short due to linguistic and contextual limitations.
The model’s release also comes at a time of heightened emphasis on AI sovereignty. With escalating geopolitical tensions and export controls limiting China’s access to the most advanced chips and cloud services, domestic AI companies face pressure to reduce dependency on foreign technologies. MM-ReasONE, optimized to run efficiently on Huawei Ascend processors and local GPU clusters, supports this goal by demonstrating that competitive reasoning AI can be achieved with largely domestic infrastructure.
In sum, MM-ReasONE is more than a commercial product; it is a strategic asset that advances China’s national AI ambitions. It reflects how closely aligned leading Chinese AI firms have become with state-driven objectives—a symbiosis that will continue to shape the evolution of AI capabilities in China for the foreseeable future.
Deep Dive into MiniMax’s New Reasoning Model
MiniMax’s latest reasoning model, officially introduced under the name “MM-ReasONE”, represents a significant leap forward in the company’s pursuit of advanced cognitive AI. The model is the product of two years of concentrated R&D, drawing from breakthroughs in transformer-based architectures, reinforcement learning with human feedback (RLHF), and a novel method of structured pretraining aimed specifically at enhancing logical reasoning capabilities.
At its core, MM-ReasONE differs from conventional large language models in that it was explicitly optimized for multi-hop reasoning, causal inference, and step-by-step problem solving rather than broad conversational fluency alone. While many LLMs have demonstrated a surface-level ability to answer factual questions or generate coherent prose, they often falter when faced with tasks requiring deep logical structure or the integration of multiple knowledge sources over extended context windows. MiniMax sought to close this gap by introducing several key innovations in its training process.
First, the model architecture incorporates dense-retrieval augmented attention mechanisms that allow MM-ReasONE to dynamically access external knowledge bases during inference. This architecture is inspired by developments in retrieval-augmented generation (RAG), but MiniMax’s engineers have refined the process to optimize for reasoning accuracy rather than simple fact retrieval. Through these mechanisms, the model can consult structured knowledge graphs and curated datasets, grounding its reasoning outputs with verifiable information.
Second, MM-ReasONE employs a multi-phase training regimen. The initial pretraining phase utilized over 5 trillion tokens, including scientific texts, legal documents, technical manuals, and mathematical problem sets—datasets that emphasize structured reasoning over casual dialogue. Following this, the model underwent supervised fine-tuning on a curated reasoning corpus derived from competition benchmarks such as Big-Bench Hard (BBH), GSM8K, and newly designed proprietary Chinese-language reasoning tests. This phase was critical in instilling the model with domain-specific reasoning capabilities in Chinese legal, financial, and scientific contexts.
A third innovation is MiniMax’s adoption of hierarchical memory layers, allowing MM-ReasONE to sustain multi-turn reasoning across extended context lengths—up to 256K tokens in certain configurations. This extended memory is particularly beneficial for applications in contract analysis, regulatory compliance, and research synthesis, where a model must reference and correlate information across large, complex documents.
Benchmark results highlight the model’s strengths. On the BBH benchmark, MM-ReasONE achieved an accuracy rate of 77.6%, surpassing DeepSeek’s 74.2%. On the GSM8K Chinese translation benchmark, designed to evaluate mathematical reasoning in Chinese, MM-ReasONE scored 83.5% accuracy—again outpacing DeepSeek and competitive with Baidu ERNIE 4.0. Notably, MM-ReasONE also performed exceptionally on the CSQA (CommonsenseQA) Chinese adaptation, registering a 90.2% correct rate, which suggests that the model excels not only at structured logic but also at commonsense reasoning—an area traditionally challenging for AI.
The model’s size is considerable, at 75 billion parameters, putting it in the same tier as DeepSeek-67B and slightly below Baidu ERNIE 4.0’s 100B configuration. However, MiniMax’s optimized architecture enables superior inference efficiency: according to internal reports, MM-ReasONE achieves 1.5x faster reasoning per token compared to DeepSeek on comparable hardware.
Training the model required access to extensive compute infrastructure. MiniMax partnered with leading Chinese cloud providers and leveraged GPU clusters based on NVIDIA H800 and Huawei Ascend 910B processors. Additionally, MiniMax developed proprietary distributed training algorithms that reduced time-to-convergence by 22% relative to its previous LLM iterations.
Finally, the model’s reinforcement learning stage is where MM-ReasONE differentiates itself most. Using RLHF techniques, human experts from fields such as law, finance, and scientific research scored and guided the model’s outputs through thousands of iterations. This human-in-the-loop training ensured that MM-ReasONE not only performed well on benchmarks but also delivered actionable, trustworthy reasoning in real-world applications.
Through these combined innovations, MiniMax has positioned MM-ReasONE as a next-generation reasoning AI model that not only challenges the current market leader DeepSeek but also elevates the overall expectations for what Chinese-developed LLMs can achieve in reasoning-intensive tasks.
Collaboration, Open Research, and the MiniMax Ecosystem
One of the distinguishing features of MiniMax’s approach to AI development is its strong emphasis on collaborative research and building an open, dynamic ecosystem. Unlike some AI companies that operate in closed silos, MiniMax has actively pursued partnerships with universities, research institutes, and government bodies. This collaborative model has not only accelerated innovation but has also helped the company attract top-tier talent and establish credibility within the Chinese and international AI research communities.
Central to this strategy is MiniMax’s involvement in joint academic-industry programs with leading institutions such as Tsinghua University, Shanghai Jiao Tong University, and the Chinese Academy of Sciences. Through these partnerships, MiniMax has contributed to open research on reasoning benchmarks, training methodologies, and ethical AI practices. The company frequently publishes peer-reviewed papers and participates in leading conferences such as ACL, NeurIPS, and EMNLP, signaling a commitment to advancing AI knowledge more broadly.
In addition, MiniMax has fostered an active developer ecosystem through its API offerings. The release of MM-ReasONE has been accompanied by comprehensive documentation, developer toolkits, and participation in open-source initiatives. This has enabled a growing community of Chinese startups, SMEs, and academic researchers to experiment with and deploy advanced reasoning AI in diverse applications. Such openness is particularly important in China’s AI ecosystem, where the ability to scale adoption beyond a handful of large tech firms is key to national innovation goals.
The company’s transparent approach has also positioned it as a trusted partner for enterprise clients and regulators. By sharing benchmark results and engaging with ethical review boards, MiniMax has built a reputation for responsible AI development—a quality that increasingly influences procurement decisions in sensitive industries. In doing so, MiniMax contributes not only to technological progress but also to shaping the governance norms of China’s AI sector.
Looking ahead, MiniMax’s collaborative ecosystem will likely serve as a platform for continued advancements in reasoning AI. As more academic partners and enterprise clients feed new data and use cases into the model’s ongoing training cycles, MM-ReasONE and its successors are poised to achieve ever higher levels of sophistication. In this sense, MiniMax is helping to define a new paradigm for AI innovation in China—one based on openness, shared knowledge, and deep integration with the nation’s broader research and industrial landscape.
Outperforming DeepSeek — What the Data Shows
MiniMax’s assertion that MM-ReasONE surpasses DeepSeek in reasoning performance is supported by a growing body of empirical data. Over the past six months, MiniMax has conducted an extensive battery of internal tests and independent evaluations, many of which have been shared with academic collaborators and government AI forums. These results suggest that MM-ReasONE not only closes the performance gap with DeepSeek but surpasses it across a range of critical reasoning benchmarks.
At the heart of this evaluation is the Big-Bench Hard (BBH) suite, an international standard for measuring the reasoning ability of advanced AI models. On BBH, MM-ReasONE achieved an overall accuracy of 77.6%, exceeding DeepSeek’s best-reported performance of 74.2%. The most notable gains were observed in tasks requiring multi-hop logical reasoning, such as legal case analysis and scientific deduction. For instance, on multi-step problem-solving tasks drawn from Chinese civil law cases, MM-ReasONE produced correct legal outcomes 12% more often than DeepSeek. These findings underscore MM-ReasONE’s superior ability to maintain logical coherence over extended chains of reasoning.
Another key benchmark is GSM8K, a dataset focused on mathematical and numerical reasoning. In its Chinese-language adaptation, MM-ReasONE scored 83.5%, compared to DeepSeek’s 79.1%. The significance of this result lies in the fact that mathematical reasoning requires precise, unambiguous logic—an area where many LLMs struggle, even with RLHF. MiniMax’s performance here points to the efficacy of its fine-tuning process and hierarchical memory architecture, which helps the model manage intermediate computations more reliably than its competitors.
The CSQA (CommonsenseQA) Chinese adaptation further highlights MM-ReasONE’s broader cognitive abilities. With a correct rate of 90.2%, MM-ReasONE demonstrates not only advanced formal reasoning but also a refined understanding of everyday human logic—an essential quality for AI systems that interact with users in real-world scenarios. DeepSeek’s score on this benchmark was 87.0%, reflecting a notable, albeit smaller, gap in commonsense reasoning performance.
In addition to structured benchmarks, MiniMax conducted head-to-head human evaluation trials involving domain experts in law, finance, and technical writing. In blind tests where experts were asked to rate the accuracy and usefulness of model-generated responses to complex queries, MM-ReasONE outperformed DeepSeek in 68% of cases. Evaluators highlighted MM-ReasONE’s superior clarity of reasoning, ability to cite relevant data, and consistency in maintaining factual accuracy across multi-turn dialogues.
One particularly illustrative example is the model’s performance on legal contract analysis—a task of significant commercial value in China’s fast-growing legal tech market. When tested on a proprietary dataset of corporate contracts, MM-ReasONE was able to correctly identify risk clauses and compliance issues with a 92% accuracy rate, compared to DeepSeek’s 86%. The model also demonstrated a more nuanced understanding of context-specific legal language, an area where many LLMs tend to underperform.
Performance is not the only domain where MM-ReasONE distinguishes itself. Inference efficiency is another critical factor, particularly for enterprise customers operating at scale. On equivalent hardware, MM-ReasONE processes reasoning-intensive queries at 1.5x the speed of DeepSeek. This advantage arises from MiniMax’s optimized attention mechanisms and improved distributed inference architecture. For applications requiring real-time or near-real-time responses—such as financial fraud detection or customer service chatbots—this speed differential translates into substantial operational benefits.
While these results collectively demonstrate MM-ReasONE’s superiority in reasoning, it is worth acknowledging that the model is not without limitations. In highly creative or open-ended generative tasks, DeepSeek and some multimodal competitors (such as SenseTime’s LLaMA-based models) continue to offer more stylistically diverse and human-like outputs. Furthermore, MM-ReasONE’s resource demands remain high, limiting its accessibility for smaller organizations.
Nevertheless, in the domain of advanced reasoning—a priority area for China’s national AI strategy—MiniMax’s latest model now holds a clear edge over DeepSeek. As independent validations continue to accumulate, MM-ReasONE’s breakthrough is poised to reshape competitive dynamics within the Chinese AI ecosystem and may spur a new wave of model development across the industry.
Industry and Strategic Implications
The unveiling of MiniMax’s MM-ReasONE model marks a pivotal moment not only for the company itself but also for China’s broader artificial intelligence industry. Its demonstrated superiority over DeepSeek on key reasoning benchmarks signals that Chinese firms are now capable of producing world-class AI systems with advanced cognitive capabilities—a domain previously dominated by a handful of US-based technology companies. The industrial and strategic ramifications of this development are wide-ranging, touching on enterprise adoption, international competitiveness, national security, and future AI innovation trajectories.
From an enterprise adoption perspective, MM-ReasONE’s enhanced reasoning abilities are already generating strong interest across multiple sectors of China’s digital economy. Financial services firms are particularly drawn to the model’s ability to conduct complex risk assessments, detect anomalous patterns, and generate regulatory-compliant reporting. Legal tech companies view MM-ReasONE as a tool that can automate contract analysis and litigation preparation with a level of nuance and accuracy previously unattainable from domestic AI models. In manufacturing and industrial domains, MM-ReasONE’s multi-step reasoning supports advanced predictive maintenance systems and optimization of complex supply chains, directly aligning with Beijing’s “smart manufacturing” initiative under Made in China 2025.
Additionally, the model’s superior inference speed makes it a practical choice for real-time applications in e-commerce, healthcare, and smart city deployments—an important consideration in China’s fast-moving consumer markets. For government agencies, the model’s capacity to interpret regulatory frameworks and provide legal-compliance guidance is being explored for public policy administration and judicial AI assistance.
At the level of international competitiveness, MM-ReasONE’s release adds to the growing evidence that Chinese AI labs can compete directly with the best Western-developed models, not only in language understanding but also in cognitive reasoning—a higher-order function critical to advanced AI. While US firms such as OpenAI (GPT-4o) and Anthropic (Claude 3) still lead in multimodal and multilingual generalization, MiniMax’s progress narrows the gap in logical reasoning and domain-specific expertise. This strengthens China’s position in the emerging global AI economy and supports Beijing’s ambitions to reduce dependency on foreign technology.
From a strategic standpoint, MM-ReasONE addresses several national priorities articulated in China’s AI development plans. The ability to create sovereign AI reasoning models that excel in law, finance, and science serves the broader objective of technological self-sufficiency—a goal further underscored by recent geopolitical tensions and US-led export controls on advanced semiconductors. Moreover, reasoning AI is increasingly recognized as a key enabler for future autonomous decision-making systems in both civilian and military contexts. MM-ReasONE’s demonstrated capacity in structured problem-solving lays the groundwork for integration into these emerging domains.
However, the release of MM-ReasONE also raises competitive pressures for other Chinese AI firms. DeepSeek, Baidu, Alibaba DAMO, and Tencent AI Lab will be compelled to accelerate their own reasoning model development, potentially fueling a new “AI reasoning arms race” within China’s tech industry. This competition is likely to intensify the demand for computational resources, drive innovation in more efficient model architectures, and stimulate greater collaboration between corporate labs and academic institutions.
There are also policy implications to consider. As reasoning models become more capable of influencing legal, financial, and governmental processes, questions about AI accountability, transparency, and ethical use will take on greater urgency. MiniMax’s relatively open approach—publishing benchmark results and engaging with academic reviewers—sets a positive example, but broader regulatory frameworks will be needed to govern the responsible deployment of such powerful systems across China’s economy.
Finally, MM-ReasONE’s success highlights several challenges that must still be addressed. Despite its strengths, the model remains compute-intensive, with substantial hardware requirements that may limit accessibility for small- and medium-sized enterprises. Talent shortages in AI reasoning research also remain a bottleneck, as demand for expert engineers continues to outpace supply. Moreover, while MM-ReasONE excels in Chinese-language tasks, its multilingual reasoning capacity is still limited compared to Western models—a gap that MiniMax must prioritize if it seeks to compete on a truly global scale.
In sum, the release of MM-ReasONE marks a strategic inflection point in China’s AI reasoning capabilities. It validates MiniMax’s position as a national leader in cognitive AI, reshapes competitive dynamics within the Chinese AI sector, and contributes meaningfully to the global evolution of reasoning-based artificial intelligence. The ripple effects of this breakthrough will be felt across industries and national AI policies in the months and years ahead.
Conclusion
The debut of MiniMax’s MM-ReasONE reasoning model represents more than a technological achievement—it is a strategic milestone that signals a maturing phase in China’s AI industry. Where once Chinese firms were primarily focused on matching Western capabilities in large language models, they are now pioneering innovations in reasoning architectures, specialized training regimes, and enterprise-grade cognitive AI systems. MM-ReasONE’s demonstrated superiority over DeepSeek in key benchmarks reflects this evolution and illustrates the increasingly sophisticated nature of domestic AI competition.
For MiniMax, the new model firmly positions the company as a top-tier player in the reasoning AI market, capable of rivaling not only national competitors but also global leaders in advanced cognitive AI. The model’s strengths—in logical reasoning, commonsense understanding, domain-specific expertise, and inference efficiency—make it highly attractive to enterprise customers across a broad spectrum of industries. From legal and financial services to manufacturing and public governance, MM-ReasONE offers tangible value and immediate applicability.
At the same time, the model’s release is poised to reshape China’s AI ecosystem. Competing firms will be under pressure to match or exceed MiniMax’s advancements, accelerating the pace of innovation and driving new investment into reasoning AI. This dynamic will likely contribute to further fragmentation of the market, as specialized models emerge to address particular industries or cognitive domains.
The implications extend beyond China’s borders. As Chinese AI firms grow increasingly competitive in reasoning-based AI, the global balance of power in artificial intelligence could shift. MiniMax’s progress also demonstrates that sovereign AI development strategies, backed by state support and robust private-sector innovation, can yield world-class results. For policymakers and industry leaders worldwide, this is a development that warrants close attention.
Looking ahead, several questions remain open. How will MiniMax address current limitations in multilingual reasoning and compute accessibility? Can the company continue to scale its capabilities in alignment with China’s evolving AI regulatory framework? And most importantly, how will MM-ReasONE influence the next generation of AI applications across sectors critical to China’s economic and strategic priorities?
What is clear is that the unveiling of MM-ReasONE marks a pivotal step in the ongoing evolution of reasoning AI. It highlights the growing technical maturity of China’s AI sector, sets new benchmarks for model performance, and reinforces the central role of advanced reasoning capabilities in shaping the future of artificial intelligence. As global competition in AI continues to intensify, developments like these will undoubtedly shape the contours of the industry for years to come.