The latest models from Chinese AI startup MiniMax are said to be on par with the best available
Chinese enterprises continue to create AI models that compete with systems built by OpenAI and other AI startups in the United States.
MiniMax, an Alibaba- and Tencent-backed business valued at more than $2.5 billion, introduced three new models this week: MiniMax-Text-01, MiniMax-VL-01, and T2A-01-HD. The company has raised around $850 million in venture funding. MiniMax-Text-01 is a text-only model, whereas MiniMax-VL-01 understands both pictures and text. Meanwhile, the T2A-01-HD creates audio, especially speech.
MiniMax says that MiniMax-Text-01, which has 456 billion parameters, outperforms models like Google's newly announced Gemini 2.0 Flash on benchmarks such as MMLU and SimpleQA, which assess a model's ability to solve math problems and fact-based questions. Parameters approximately equate to a model's problem-solving abilities, and models with more parameters typically outperform those with fewer parameters.
MiniMax-VL-01 is said to compete with Anthropic's Claude 3.5 Sonnet on assessments that demand multimodal comprehension, such as ChartQA, which challenges models with answering graph- and diagram-related inquiries (e.g., "What is the peak value of the orange line in this graph?"). Granted, MiniMax-VL-01 does not quite outperform Gemini 2.0 Flash in many of these tests. OpenAI's GPT-4o and an open model called InternVL2.5 also outperformed it in numerous areas.
MiniMax-Text-01 features a very wide context window. A model's context, also known as the context window, refers to the input (for example, text) that a model takes into account before producing output (more text). With a context window of 4 million tokens, MiniMax-Text-01 can analyze around 3 million words at once, or a little more than five copies of "War and Peace."
MiniMax-Text-01's context window is approximately 31 times larger than GPT-4o's and Llama 3.1's.
T2A-01-HD, the final of MiniMax's products introduced this week, is an audio generator designed for speech. The T2A-01-HD can manufacture a synthetic voice with changeable cadence, tone, and tenor in about 17 languages, including English and Chinese, as well as clone a voice from a 10-second audio recording.
MiniMax did not disclose benchmark data comparing the T2A-01-HD to comparable audio-generating devices. However, to this reporter's ears, the T2A-01-HD's outputs sound comparable to audio models from Meta and startups such as PlayAI.
With the exception of T2A-01-HD, which is only available through MiniMax's API and the Hailuo AI platform, MiniMax's latest models are available for download on GitHub and the AI development platform Hugging Face.
The fact that the models are "openly" available does not imply that they are not restricted in any way. MiniMax-Text-01 and MiniMax-VL-01 are not genuinely open source in the sense that MiniMax has not made available the components (such as training data) required to recreate them from scratch. Furthermore, they are subject to MiniMax's restrictive licensing, which prevents developers from utilizing the models to enhance competing AI models and requires platforms with more than 100 million monthly active users to seek a special license from MiniMax.
MiniMax was created in 2021 by former workers of SenseTime, China's top AI company. MiniMax's efforts include Talkie, an AI-powered role-playing platform similar to Character AI, and text-to-video models released in Hawaii.
Some MiniMax items have sparked little controversy.
Talkie, which was removed from Apple's App Store in December for vague "technical" reasons, includes AI avatars of popular celebrities such as Donald Trump, Taylor Swift, Elon Musk, and LeBron James, none of whom appear to have agreed to be included in the app.
Broadcast magazine stated in December that MiniMax's video generators can imitate the logos of British television networks, implying that the models were trained on that programming. MiniMax is also apparently being sued by iQiyi, a Chinese video streaming service, which claims MiniMax illegally trained on its copyrighted recordings.
MiniMax's latest models came only days after the departing Biden administration suggested stricter export controls and limits on AI technologies for Chinese companies. They were previously prohibited from purchasing powerful AI chips, but if the new laws are implemented as intended, they would face tougher limits on both semiconductor technology and models required to bootstrap complex AI systems.
On Wednesday, the Biden administration unveiled more safeguards to keep advanced electronics out of China. Chip foundries and packaging businesses that wish to export specific chips will face more stringent licensing requirements unless they take greater caution and due diligence to prevent their goods from reaching Chinese customers.