China AI rising: Xiaomi releases new MiMo-7B fashions as DeepSeek upgrades its Prover math AI

Xiaomi Corp. as we speak launched MiMo-7B, a brand new household of reasoning fashions that it claims can outperform OpenAI’s o1-mini at some duties.

The algorithm sequence is accessible below an open-source license. Its launch coincides with DeepSeek’s launch of an replace to Prover, a competing open-source reasoning mannequin. The latter algorithm has a narrower focus than MiMo-7B: it’s designed to assist mathematicians show theorems.

The algorithms in Xiaomi’s MiMo-7B sequence have about 7 billion parameters. There’s a base mannequin, in addition to enhanced variations of that mannequin that supply elevated output high quality.

Xiaomi developed the improved variations utilizing two machine studying methods known as supervised fine-tuning and reinforcement studying. Each strategies enhance AI fashions by offering them with extra coaching information. The datasets utilized in supervised fine-tuning embrace explainers that assist information the AI coaching workflow, whereas reinforcement studying doesn’t use such explainers.

The corporate has developed three enhanced variations of the MiMo-7B base mannequin. It fined-tuned one model utilizing supervised fine-tuning, one other with reinforcement studying, and a 3rd utilizing each strategies. In line with the corporate, that third mannequin is healthier than OpenAI’s o1-mini at producing code and fixing math issues.

The bottom MiMo-7B mannequin is much less succesful than the fine-tuned variations, however can nonetheless outdo considerably bigger algorithms. “Our RL experiments from MiMo-7B-Base present that our mannequin possesses extraordinary reasoning potential, even surpassing a lot bigger 32B fashions,” Xiaomi researchers detailed on GitHub.

The MiMo-7B sequence shouldn’t be the one new entry into the open-source AI ecosystem that debuted as we speak. DeepSeek quietly launched an enhanced model of Prover, a reasoning mannequin optimized to show mathematical theorems that it first debuted final yr. Prover-V2, because the upgraded mannequin is named, guarantees to offer “state-of-the-art efficiency in neural theorem proving.”

DeepSeek educated Prover-V2 by means of a multistep course of. The corporate began by assembling a set of theorems for which proofs are already out there. Within the subsequent step, DeepSeek used two language fashions to create a step-by-step clarification of how mathematicians arrived at every proof. The corporate subsequently entered these AI-generation explanations into Prover V2 to show the mannequin tips on how to generate its personal proofs.

“This course of allows us to combine each casual and formal mathematical reasoning right into a unified mannequin,” DeepSeek researchers defined. 

The discharge of MiMo-7B and Prover-V2 comes days after Alibaba Group Holding Ltd. launched Qwen3, its new flagship household of reasoning-optimized fashions. The algorithms within the sequence vary in measurement from 600 million to 235 billion parameters. Alibaba claims that Qwen3 can outperform OpenAI’s o1 and DeepSeek’s flagship R1 reasoning mannequin throughout a variety of duties. 

Picture: Unsplash