Chinese language corporations proceed to launch AI fashions that rival the capabilities of methods developed by OpenAI and different U.S.-based AI firms.
This week, MiniMax, an Alibaba- and Tencent-backed startup that has raised round $850 million in enterprise capital and is valued at greater than $2.5 billion, debuted three new fashions: MiniMax-Textual content-01, MiniMax-VL-01, and T2A-01-HD. MiniMax-Textual content-01 is a text-only mannequin, whereas MiniMax-VL-01 can perceive each pictures and textual content. T2A-01-HD, in the meantime, generates audio — particularly speech.
MiniMax claims that MiniMax-Textual content-01, which is 456 billion parameters in dimension, performs higher than fashions comparable to Google’s just lately unveiled Gemini 2.0 Flash on benchmarks like MMLU and SimpleQA, which measure the flexibility of a mannequin to reply math issues and fact-based questions. Parameters roughly correspond to a mannequin’s problem-solving abilities, and fashions with extra parameters usually carry out higher than these with fewer parameters.
As for MiniMax-VL-01, MiniMax says that it rivals Anthropic’s Claude 3.5 Sonnet on evaluations that require multimodal understanding, like ChartQA, which duties fashions with answering graph- and diagram-related queries (e.g., “What’s the peak worth of the orange line on this graph?”). Granted, MiniMax-VL-01 doesn’t fairly greatest Gemini 2.0 Flash on many of those assessments. OpenAI’s GPT-4o and an open mannequin referred to as InternVL2.5 beat it on a number of as nicely.
Of observe, MiniMax-Textual content-01 has a particularly massive context window. A mannequin’s context, or context window, refers to enter (e.g., textual content) {that a} mannequin considers earlier than producing output (further textual content). With a context window of 4 million tokens, MiniMax-Textual content-01 can analyze round 3 million phrases in a single go — or simply over 5 copies of “Conflict and Peace.”
For context (no pun supposed), MiniMax-Textual content-01’s context window is roughly 31 instances the dimensions of GPT-4o’s and Llama 3.1’s.
The final of MiniMax’s fashions launched this week, T2A-01-HD, is an audio generator optimized for speech. T2A-01-HD can generate an artificial voice with adjustable cadence, tone, and tenor in round 17 completely different languages, together with English and Chinese language, and clone a voice from simply 10 seconds of an audio recording.
MiniMax didn’t publish benchmark outcomes evaluating T2A-01-HD to different audio-generating fashions. However to this reporter’s ear, T2A-01-HD’s outputs sound on par with audio fashions from Meta and startups like PlayAI.
Except for T2A-01-HD, which is solely obtainable by means of MiniMax’s API and Hailuo AI platform, MiniMax’s new fashions might be downloaded from GitHub and the AI dev platform Hugging Face.
Simply because the fashions are “brazenly” obtainable doesn’t imply they aren’t locked down in sure facets, nonetheless. MiniMax-Textual content-01 and MiniMax-VL-01 aren’t actually open supply within the sense that MiniMax hasn’t launched the elements (e.g., coaching information) wanted to re-create them from scratch. Furthermore, they’re below MiniMax’s restrictive license, which prohibits builders from utilizing the fashions to enhance rival AI fashions and requires that platforms with greater than 100 million month-to-month lively customers request a particular license from MiniMax.
MiniMax was based in 2021 by former staff of SenseTime, one in every of China’s largest AI corporations. The corporate’s initiatives embrace apps like Talkie, an AI-powered role-playing platform alongside the strains of Character AI, and text-to-video fashions that MiniMax has launched in Hailuo.
A few of MiniMax’s merchandise have develop into the topic of minor controversy.
Talkie, which was pulled from Apple’s App Retailer in December for unspecified “technical” causes, options AI avatars of public figures, together with Donald Trump, Taylor Swift, Elon Musk, and LeBron James, none of whom seem to have consented to being featured within the app.
In December, Broadcast journal reported that MiniMax’s video mills can reproduce the logos of British tv channels, suggesting that MiniMax’s fashions have been educated on content material from these channels. And MiniMax is reportedly being sued by iQiyi, a Chinese language video streaming service that alleges MiniMax illicitly educated on iQiyi’s copyrighted recordings.
MiniMax’s new fashions arrive days after the outgoing Biden administration proposed harsher export guidelines and restrictions on AI applied sciences for Chinese language ventures. Corporations in China have been already prevented from shopping for superior AI chips, but when the brand new guidelines go into impact as written, firms shall be confronted with stricter caps on each the semiconductor tech and fashions wanted to bootstrap subtle AI methods.
On Wednesday, the Biden administration introduced further measures targeted on retaining subtle chips out of China. Chip foundries and packaging firms that need to export sure chips shall be subjected to broader license necessities until they train better scrutiny and due diligence to stop their merchandise from reaching Chinese language purchasers.
