Open-source AI model developer MiniMax raises $2B in funding

MiniMax Group Inc., a Shanghai-based artificial intelligence developer, is raising $2 billion in funding.

Bloomberg reported on Thursday that more than half the funds are expected to come from the sale of newly issued shares. According to the publication, the company will raise the remaining capital by issuing convertible bonds. Those are debt instruments that investors can turn into stock.

MiniMax reportedly plans to follow up the raise by selling $6.5 billion worth of zero-coupon convertible bonds, which are convertible bonds that don’t generate interest. They will enable investors to obtain 19.4 million shares priced 12.6% higher than the company’s Thursday closing price. The bonds are reportedly due in 2027.

Shares of MiniMax closed 9.8% lower on the news today. The company listed its shares in Hong Kong earlier this year through a public offering that raised about $619 million.

MiniMax develops an eponymous series of open-source large language models. The company debuted its newest algorithm, MiniMax-M3, in June. It features 427 billion parameters and can process prompts that contain up to 1 million tokens.

The process through which an LLM answers a prompt comprises two main steps known as the prefill and decode phases. MiniMax says M3 can perform those steps nine and 15 times faster, respectively, than its previous flagship LLM. The company credits the speedup to a technology it calls MiniMax Sparse Attention, or MSA.

Graphics cards include two types of memory known as SRAM and HBM. LLMs regularly move data between them while performing inference. MiniMax’s MSA technology uses a machine learning learning approach called FlashAttention to reduce the number of times an LLM shifts data between SRAM and HBM, which speeds up inference.

MSA also optimizes model performance in other ways. It uses a technique known as block-sparse prefill to help M3 more quickly understand large prompts. Additionally, a quantization module compresses some of the data that the LLM generates to reduce its memory requirements.

MiniMax’s research focus also extends beyond LLMs. It has developed a family of open-source visual tokenizers known as the VTL series. A visualizer tokenizer is an AI model that turns images into vectors, abstract mathematical representations that LLMs can more easily process.

MiniMax generates revenue with a cloud-based platform that provides access to hosted versions of its models. Additionally, the company offers paid multimedia generation apps geared towards consumers.

The Information reported on Thursday that MiniMax Chief Executive Officer Yan Junjie had announced plans to forgo his salary until the company develops artificial general intelligence, or AGI. He also pledged to give away 5% of his shares. The equity will be used to finance employee incentives and open-source projects.

Image: Pixabay