China’s MiniMax launches M1: A reasoning model to rival GPT-4 at 0.5% the cost

Chinese AI startup MiniMax has thrown down the gauntlet to established AI giants, releasing what it boldly claims is the world’s first open-source, large-scale hybrid-attention reasoning model that could fundamentally change the economics of advanced AI development.

MiniMax defines “hybrid-attention” as a combination of its Lightning Attention mechanism and Mixture-of-Experts architecture, which activates only the relevant parts of the model for each task.

The Shanghai-based company said its new MiniMax-M1 model delivers a knockout punch to computational inefficiency, requiring just 30% of the computing power needed by rival DeepSeek’s R1 model when performing deep reasoning tasks — a breakthrough that could democratize access to sophisticated AI capabilities.

“In complex, productivity-oriented scenarios, M1’s capabilities are top-tier among open-source models, surpassing domestic closed-source models and approaching the leading overseas models, all while offering the industry’s best cost-effectiveness,” the company said, announcing the hybrid-attention reasoning model.

A David vs. Goliath moment in AI

What makes MiniMax-M1 particularly intriguing isn’t just its performance — it’s how the company achieved it. While tech titans have been throwing hundreds of millions of dollars at AI development, MiniMax managed to train its model for a mere $534,700.

To put that in perspective, DeepSeek spent $5-$6 million on its R1 model, while OpenAI’s GPT-4 reportedly cost over $100 million to develop.

“The entire reinforcement learning phase used only 512 H800s for three weeks, with a rental cost of just $534,700,” the company explained. “This is an order of magnitude less than initially anticipated.”

However, industry analysts urge caution. “MiniMax’s debut reasoning model, M1, has generated justified excitement with its claim of reducing computational demands by up to 70% compared to peers like DeepSeek-R1,” said Sanchit Vir Gogia, chief analyst and CEO at Greyhound Research. “However, amid growing scrutiny of AI benchmarking practices, enterprises must independently replicate such claims across practical workloads.”

Extended context capabilities

M1’s most impressive feature might be its ability to handle massive amounts of information simultaneously. With a context window supporting one million input tokens and up to 80,000 output tokens, the model can essentially read and understand multiple novel tasks at once while maintaining coherence across the entire conversation.

“A significant advantage of M1 is its support for an industry-leading 1 million token context window, matching the closed-source Google Gemini 2.5 Pro,” MiniMax noted in the post. “This is 8 times that of DeepSeek R1 and includes an industry-leading 80,000 token reasoning output.”

For context, OpenAI’s GPT-4o can handle only 128,000 tokens — enough for about one novel task. M1’s expanded capacity opens doors for applications that were previously impractical, from analyzing entire legal documents to debugging massive code repositories.

Real-world performance that matters

Beyond impressive technical specifications, M1 demonstrates strong real-world capabilities across multiple benchmarks. The model comes in two variants — M1-40k and M1-80k, referring to their “thinking budgets” — with the larger version consistently outperforming its smaller sibling across most tests.

In software engineering tasks, both versions scored 55.6% and 56.0%, respectively, on the challenging SWE-bench validation benchmark. While slightly trailing DeepSeek-R1-0528’s 57.6%, they significantly outpaced other open-weight models in this critical area of productivity.

“MiniMax-M1-80k consistently outperforms MiniMax-M1-40k across most benchmarks, which fully validates the effectiveness of extended computational resources during testing,” the company added.

Breaking down barriers to AI access

Perhaps most significantly, MiniMax is releasing M1 under a true Apache 2.0 open-source license — unlike Meta’s Llama models, which use restrictive community licenses, or DeepSeek’s partially open approach. This decision could accelerate innovation by giving researchers and developers unprecedented access to cutting-edge reasoning capabilities.

Gogia sees this as particularly significant for mid-market companies. “MiniMax’s M1 represents more than just architectural efficiency — it symbolizes the new accessibility of advanced reasoning AI for mid-market enterprises,” he noted. “With open-source licensing, reduced inference costs, and support for 1 M-token context windows, M1 aligns squarely with the evolving needs of midsize firms that seek capability parity with larger players but lack hyperscaler budgets.”

The company is backing up its open-source commitment with competitive pricing for those who prefer API access. Input processing costs just $0.4 per million tokens for contexts up to 200,000 tokens, rising to $1.3 per million tokens for the full 1-million-token capability.

“Due to its relatively efficient use of training and inference computing power, we are offering unlimited free use on the MiniMax APP and Web, and providing APIs on our official website at the industry’s lowest prices,” the company announced.

What this means for the industry

M1’s release comes at a pivotal moment in AI development, as the industry grapples with the massive computational costs of training and running advanced models.

The timing is particularly noteworthy, coming just weeks after fellow Chinese company DeepSeek shook the industry with its own cost-effective approach to AI development. This pattern suggests Chinese companies are finding innovative ways to compete with better-funded Western rivals through superior engineering rather than just throwing money at problems.

Yet challenges remain for Chinese AI models in Western markets. Despite technical achievements, Gogia notes that “Chinese LLMs remain under-adopted in North America and Western Europe” due to concerns around governance and regulatory compliance in industries with strict procurement frameworks.

A company on the rise

MiniMax isn’t exactly a household name yet, but it’s rapidly becoming one of China’s most-watched AI companies. Founded in 2021 by former SenseTime executives, the startup has raised $850 million from heavyweight investors including Alibaba, Tencent, and IDG Capital, achieving a $2.5 billion valuation.

This M1 announcement kicks off what MiniMax is calling “MiniMaxWeek,” with additional product releases planned for the coming days. For enterprise users and developers, M1 represents something potentially transformative: enterprise-grade AI reasoning capabilities without enterprise-grade infrastructure requirements. However, as Gogia cautioned, “The real test will lie in how quickly CIOs can extract operational savings at scale, without compromising accuracy or governance.”