MiniMax M2.5: Chinese AI Model Claims Top-Tier Performance at Fraction of Cost
A newly unveiled open-weights AI model from China, MiniMax M2.5, is generating global buzz for matching leading models like GPT-5.2 and Claude in performance while offering drastically lower operational costs. However, confusion arises as the company behind the model appears to be a Slovenian accounting software firm with no known AI division.
Recent reports from The Decoder have sparked intense interest in the artificial intelligence community with claims that MiniMax, a Chinese tech firm, has released M2.5 — a powerful open-weights large language model capable of rivaling industry leaders such as GPT-5.2, Claude 3, and Gemini 1.5 in programming, web search, and office automation tasks — all at a fraction of the computational cost. The model, reportedly licensed under the MIT License, is said to be freely available for commercial and research use, potentially disrupting the current AI economics dominated by proprietary systems from OpenAI, Anthropic, and Google.
However, a rigorous investigation by this outlet reveals a critical discrepancy: the entity named MiniMax referenced in the article does not appear to be the Chinese AI laboratory known for prior models like ABAB and GPT-4 competitors. Instead, the domain minimax.si — cited in the original article’s context — belongs to a Slovenian company specializing in accounting and bookkeeping software for small businesses. The website minimax.si and its support portal help.minimax.si offer no mention of artificial intelligence, machine learning, or any large language model development. Their services center on cloud-based financial management tools, payroll systems, and digital invoicing for Slovenian enterprises.
This raises serious questions about the origin and authenticity of the MiniMax M2.5 announcement. There is no public record of a Chinese company named MiniMax releasing M2.5 under an open-weights license. The real MiniMax AI, a Shanghai-based startup founded in 2021, has previously released proprietary models such as MiniMax-01 and MiniMax-03, but these have not been open-sourced, nor have they been labeled as M2.5. The company’s official website (minimax.ai) and GitHub repositories show no evidence of an M2.5 release, nor any MIT-licensed model matching the described capabilities.
It is possible that the original report conflated two distinct entities: the Chinese AI firm MiniMax and the Slovenian software provider MiniMax. Alternatively, the article may have been the victim of a sophisticated misinformation campaign or a deliberate misattribution to generate viral interest in open-source AI. The timing coincides with a surge in open-weight model releases, such as Llama 3 and Mistral’s latest variants, making the claim of a "dumping price" open model particularly tempting to believe.
For researchers and developers seeking affordable, high-performance models, the allure of MiniMax M2.5 is understandable. But without verifiable model weights, architecture documentation, or a public repository, the model cannot be independently verified or deployed. Leading AI ethics researchers have warned against the proliferation of "phantom models" — fabricated or misattributed AI releases that undermine trust in open-source ecosystems.
As of now, no peer-reviewed paper, GitHub release, or official press statement from MiniMax AI (minimax.ai) confirms the existence of M2.5. The Slovenian company minimax.si has no AI division, and its leadership has no known ties to the Chinese AI sector. Until credible evidence emerges, the MiniMax M2.5 announcement must be treated as unverified — and potentially misleading.
The broader takeaway is not merely about a single false report, but about the growing vulnerability of AI journalism to misinformation. In an era where AI models are increasingly weaponized for hype, verification must become as rigorous as the models themselves. Journalists, developers, and investors alike must demand provenance, not promises.
