MiniMax Launches M2.5 Models with Low Cost and High Performance

The Chinese artificial intelligence startup, MiniMax, announced today the launch of the M2.5 language model in two versions, promising to democratize access to advanced artificial intelligence with significantly reduced costs. The M2.5 can cost up to 95% less compared to cutting-edge models like Claude Opus 4.6, challenging the traditional high investment in conventional AI technology.

Although the model is announced as "open source", details about weights and license terms have not yet been released. MiniMax focuses on accessibility, offering services through its API and partnerships.

The models offer capabilities that rival those of industry giants like Google and Anthropic, being particularly efficient in business tasks with office document automation, such as Word, Excel, and PowerPoint. "This launch signals a shift in perception about AI, evolving from a conversational tool to an effective work agent," said the MiniMax team.

With 30% of tasks in MiniMax’s office completed by the M2.5 model and 80% of the new code generated by it, the company demonstrates confidence in its effectiveness. According to MiniMax's blog, "the M2.5 offers unlimited possibilities for the development and operation of agents in the economy."

Technology: efficiency through MoE architecture

The efficiency of the M2.5 model is based on an architecture called Mixture of Experts (MoE). This technique allows only 10 billion of the 230 billion parameters to be activated simultaneously for generating each word, maintaining the reasoning depth of a large model while operating with the agility of a smaller model.

To train the M2.5, MiniMax developed a Reinforcement Learning (RL) framework called Forge. During the ThursdAI podcast, engineer Olive Song emphasized that this technique was crucial for maximizing performance with a smaller number of parameters.

Additionally, MiniMax utilizes a mathematical approach known as CISPO (Clipping Importance Sampling Policy Optimization) to ensure stability during intensive training, allowing the M2.5 to develop an "Architect's Mindset," learning to plan projects before coding.

Performance and comparison with leading models

The performance of the M2.5 positions it among the best in the industry. Approaching the performance of Claude Opus 4.6, the new benchmark results for the M2.5 include:

• SWE-Bench Verified: 80.2% - Comparable speeds to Claude Opus 4.6.
• BrowseComp: 76.3% - Leader in searches and tool usage.
• Multi-SWE-Bench: 51.3% - Top performance in multilingual coding.
• BFCL (Tool Call): 76.8% - High level of accuracy in workflows.

On the ThursdAI podcast, it was highlighted that the MiniMax M2.5 operates quickly, consuming fewer tokens, with a cost of only $0.15 per task compared to $3.00 for Claude Opus 4.6.

Demystifying cost barriers

MiniMax offers two versions of the M2.5 through its API:

• M2.5-Lightning: Focused on speed, with a cost of $0.30 per 1M input tokens and $2.40 per 1M output tokens.
• Standard M2.5: Focused on cost, with a value of $0.15 per 1M input tokens and $1.20 per 1M output tokens.

These values make it feasible to use four agents continuously for about $10,000 over a year, presenting a cost of 1/10 to 1/20 of competing models like GPT-5 or Claude 4.6. A detailed price comparison has been made available to contextualize this innovation.

Strategic implications for businesses

The M2.5 represents a shift in the operational model for business leaders, as it eliminates the pressure for cost optimization in processes that were previously considered unfeasible. With a 37% improvement in task completion speed, the M2.5 enables agile pipelines that facilitate effective communication between different AI models.

Moreover, the high score in financial modeling indicates that the model can handle the tacit knowledge necessary in specialized industries like finance and law. Offering the M2.5 as an open-source model enables large-scale automated audits, increasing control over data privacy while awaiting details on license conditions and weights.

The MiniMax M2.5 signals that the future of artificial intelligence is not just about building the most complex model but about who can make that model useful and accessible in the workplace.