Chinese companies continue to release AI models that rival the capabilities of systems developed by OpenAI and other US-based AI companies.

this week, MiniMaxan Alibaba- and Tencent-backed startup with raised about $850 million in venture capital and valued at over $2.5 billion, debuted leakage new models: MiniMax-Text-01, MiniMax-VL-01, and T2A-01-HD. The MiniMax-Text-01 is a text-only model, while the MiniMax-VL-01 can understand both images and text. The T2A-01-HD, on the other hand, produces audio – especially speech.

MiniMax claims that MiniMax-Text-01, with 456 billion size parameters, is better than models like the one recently unveiled by Google. Gemini 2.0 Flash in benchmarks such as MATH and SimpleQA, which measure a model’s ability to answer mathematical problems and fact-based questions. Parameters roughly correspond to a model’s problem-solving skill, and models with more parameters generally perform better than those with fewer parameters.

As for the MiniMax-VL-01, MiniMax says it’s the opposite of Anthropic’s Claude 3.5 Sonnet in evaluations that require multimodal understanding, such as ChartQA, which operates on models to answer questions related to graphs and diagrams (eg, “What is the maximum value of the orange line on this graph?” ). Admittedly, the MiniMax-VL-01 didn’t quite outperform the Gemini 2.0 Flash in most of these tests. OpenAIs GPT-4o and Meta Name 3.1 some beat it too.

Of note, MiniMax-Text-01 has a very large context window. A model’s context, or context window, refers to the input (eg, text) that a model considers before producing output (additional text). With a context window of 4 million tokens, MiniMax-Text-01 can analyze almost 3 million words in one go – or more than five copies of “War and Peace.”

For context (no pun intended), MiniMax-Text-01’s context window is roughly 31 times the size of GPT-4o’s and Llama 3.1’s.

The last of the MiniMax models released this week, the T2A-01-HD, is an audio generator optimized for speech. The T2A-01-HD can produce a synthetic voice with adjustable cadence, tone, and tenor in about 17 different languages, including English and Chinese, and clone a voice from 10 seconds just an audio recording.

MiniMax does not publish benchmark results comparing the T2A-01-HD to other audio-generating models. But to this reporter’s ears, the T2A-01-HD’s outputs sound like audio models from Meta and startups like PlayAI.

Except for the T2A-01-HD, which is exclusively available through MiniMax’s API and Hailuo AI platform, the new MiniMax models can be downloaded from GitHub and the AI ​​dev platform Hugging Face.

Just because models are “open” to use doesn’t mean they aren’t locked in certain aspects, however. MiniMax-Text-01 and MiniMax-VL-01 not really open source in the sense that MiniMax does not release the components (eg, training data) needed to recreate it from scratch. In addition, they are under the strict MiniMax license, which prohibits developers from using the models to improve rival AI models and requires that platforms with more than 100 million monthly active users users request a special license from MiniMax.

MiniMax was founded in 2021 by former employees of SenseTime, one of the largest AI companies in China. The company’s projects include apps such as Talkie, an AI-powered role-playing platform along the lines of Character AIand text-to-video models released by Hailuo’s MiniMax.

Some of MiniMax’s products have been the subject of minor controversy.

Talkie, which was pulled from Apple’s App Store in December for unspecified “technical” reasons, features AI avatars of public figures, including Donald Trump, Taylor Swift, Elon Musk, and LeBron James, that no one who appeared agreed to be shown on. the app.

In December, Broadcast magazine reported that the MiniMax video generators were able to reproduce the logos of British television channels, suggesting that the MiniMax models were trained internally from the channels. And the MiniMax is reported nabbed by iQiyi, a Chinese video streaming service that said MiniMax was illegally trained on iQiyi’s copyrighted recordings.

The new MiniMax models come days after the outgoing Biden administration proposed stricter export rules and restrictions on AI technologies for Chinese businesses. Chinese companies are already restricted from buying advanced AI chips, but if the new rules go into effect as written, companies will face stricter caps on both the semiconductor tech and models required. to bootstrap sophisticated AI systems.

On Wednesday, the Biden administration Office has partnered further steps focused on keeping sophisticated chips from China. Chip foundries and packaging companies that want to export certain chips will be subject to broader licensing requirements unless they use greater scrutiny and due diligence to prevent their products from reaching the client in China.