Chinese companies continue to release AI models that rival the capabilities of systems developed by OpenAI and other US-based AI companies.
This week, MiniMax, a startup backed by Alibaba and Tencent, raised about $850 million in venture capital at a valuation of more than $2.5 billion and has launched MiniMax-Text-01, MiniMax-VL-01, and debuted three new models of T2A. -01-HD. MiniMax-Text-01 is a text-only model, while MiniMax-VL-01 can understand both images and text. The T2A-01-HD, on the other hand, produces audio, especially audio.
With a size of 456 billion parameters, MiniMax-Text-01 outperforms models such as Google's recently announced Gemini 2.0 Flash on benchmarks such as MATH and SimpleQA, which measure a model's ability to answer math problems and facts. It claims to have excellent performance. Questions based on. Parameters roughly correspond to the model's problem-solving skills, and models with more parameters generally perform better than models with fewer parameters.
Regarding the MiniMax-VL-01, MiniMax says it is comparable to Anthropic's Claude 3.5 Sonnet in assessments that require multimodal understanding, such as ChartQA. In ChartQA, queries related to charts and diagrams (e.g. “What is the orange line on this chart?”). Admittedly, the MiniMax-VL-01 doesn't outperform the Gemini 2.0 flash in many of these tests. OpenAI's GPT-4o and Meta's Llama 3.1 also surpassed this in several ways.
One thing to note is that MiniMax-Text-01 has a very large context window. A model's context, or context window, refers to the input (such as text) that the model considers before producing output (additional text). MiniMax-Text-01 features a 4 million token context window and can analyze approximately 3 million words at a time, or just over five copies of “War and Peace.”
Speaking of context (I'm not kidding), MiniMax-Text-01's context window is about 31 times the size of GPT-4o and Llama 3.1.
MiniMax's final model, the T2A-01-HD, released this week, is a voice-optimized audio generator. The T2A-01-HD produces synthetic voices with adjustable rhythm, tone, and tenor in approximately 17 languages, including English and Chinese, and can clone voices from just 10 seconds of audio recording.
MiniMax does not publish benchmark results comparing the T2A-01-HD to other audio generation models. But to our reporters' ears, the T2A-01-HD outputs sound on par with audio models from startups like Meta and PlayAI.
With the exception of T2A-01-HD, which is available exclusively through MiniMax's API and Hailuo AI platform, MiniMax's new models can be downloaded from GitHub and the AI ​​development platform Hugging Face.
However, just because a model is “openly” available does not mean it is not locked down in certain aspects. MiniMax-Text-01 and MiniMax-VL-01 are not truly open source in the sense that MiniMax has not released the necessary components (such as training data) to recreate them from scratch. Additionally, they are under MiniMax's restrictive license, which prohibits developers from using the models to improve rival AI models, and platforms with more than 100 million monthly active users receive special licenses from MiniMax. license.
MiniMax was founded in 2021 by former employees of SenseTime, one of China's largest AI companies. The company's projects include apps like Talkie, an AI-powered role-playing platform in the vein of Character AI, and the text-to-video model MiniMax released in Haira.
Some of MiniMax's products have been the subject of minor controversy.
Talkie, which was removed from Apple's App Store in December for unspecified “technical” reasons, features AI avatars of celebrities including Donald Trump, Taylor Swift, Elon Musk and LeBron James. However, none of them seem to have consented to publication. App.
In December, Broadcast magazine reported that MiniMax's video generator could reproduce the logos of British TV channels, suggesting that MiniMax's models were trained on content from those channels. MiniMax is also reportedly being sued by iQiyi, a Chinese video streaming service, alleging that MiniMax was illegally trained using iQiyi's copyrighted recordings.
MiniMax's new model comes days after the outgoing Biden administration proposed tougher export controls and restrictions on AI technology for Chinese ventures. Chinese companies are already prohibited from purchasing advanced AI chips, but if the new rules take effect as written, companies will face tighter restrictions on both the semiconductor technology and models needed to build advanced AI systems. You will have to face it.
The Biden administration on Wednesday announced additional measures focused on keeping high-performance chips out of China. Chip foundries and packaging companies wishing to export certain chips will be subject to broader licensing requirements unless they implement greater oversight and due diligence to prevent their products from reaching Chinese customers. Probably.