[Digital Today reporter Chi-gyu Hwang (황치규)] Minimax, an artificial intelligence developer in Shanghai, China, has begun raising $2 billion. SiliconANGLE reported on Thursday that Minimax plans to raise more than half of the funds by issuing new shares, with the rest to be raised through convertible bonds.

Minimax is also pushing ahead with a $6.5 billion convertible bond offering. The bonds mature in 2027, and investors can obtain 19.4 million shares at a price 12.6 percent above Thursday’s close.

Minimax shares fell 9.8 percent on the day after the fundraising news emerged. Earlier this year, Minimax raised about $619 million through a listing on the Hong Kong stock market.

Minimax develops an open-source large language model series. Its latest model, MiniMax-M3, released in June, has 427 billion parameters and can handle prompts up to 1 million tokens in length.

Minimax said M3 is 9 times faster in the prefill stage and 15 times faster in the decode stage than its previous flagship model. The speed gains use MiniMax Sparse Attention. The technology is a FlashAttention-based technique that reduces data movement between SRAM and HBM during inference.

M3 also applies a block sparse prefill technique that processes longer inputs faster, and a quantization module that compresses some generated data to reduce memory requirements. Minimax also developed an open-source visual tokenizer VTL series that converts images into vectors to make them easier for other models to process.

Minimax earns revenue from its model hosting service and a paid consumer multimedia creation app. Minimax CEO Junjie Yan (얀쥔제) said he plans not to receive a salary until the company develops artificial general intelligence, and also said he would allocate 5 percent of his stake to employee incentives and funding for open-source projects.