Xiaomi’s MiMo AI Models Arrive With Efficient Reasoning, Small Size

Xiaomi on Tuesday launched an open-source reasoning-focused synthetic intelligence (AI) mannequin. Dubbed MiMo, the household of reasoning fashions innovate the optimisation of reasoning functionality in a comparatively smaller parameter measurement. That is additionally the primary open-source reasoning mannequin by the tech big, and it competes with Chinese language fashions similar to DeepSeek R1 and Alibaba’s Qwen QwQ-32B, and world reasoning fashions together with OpenAI’s o1 and Google’s Gemini 2.0 Flash Pondering. The MiMo household includes 4 completely different fashions, every with distinctive use instances.

Xiaomi’s MiMo Reasoning AI Mannequin to Compete With DeepSeek R1

With the MiMo collection of AI fashions, Xiaomi researchers aimed to unravel the scale drawback in reasoning AI fashions. Reasoning fashions (a minimum of ones that may be measured) have round 24 billion or extra parameters. The massive measurement is saved to attain uniform and simultaneous enhancements in each coding and mathematical capabilities of enormous language fashions, one thing thought of tough to attain with smaller fashions.

Compared, MiMo options seven billion parameters, and Xiaomi claims that its efficiency matches OpenAI’s o1-mini and outperforms a number of reasoning fashions with 32 billion parameters. The researchers claimed that the bottom AI mannequin was pre-trained on 25 trillion tokens.

The researchers claimed that such effectivity was achieved by optimising information preprocessing pipelines, enhancing textual content extraction toolkits, and making use of multidimensional information filtering. Additional, MiMo’s pre-training included a three-stage information combination technique.

Based mostly on inside testing, the Xiaomi researchers declare that the MiMo-7B-Base scores 75.2 on the BIG-Bench Laborious (BBH) benchmark for reasoning capabilities. The zero-shot reinforcement studying (RL)-based MiMo-7B-RL-Zero is claimed to excel in arithmetic and coding-related duties, and scores 55.4 on the AIME benchmark, outperforming o1-mini by 4.7 factors.

As MiMo is an open-source AI mannequin, it may be downloaded from Xiaomi’s itemizing on GitHub and Hugging Face. The technical paper particulars the mannequin’s structure in addition to the pre-training and post-training processes. It’s a text-based mannequin and doesn’t have multimodal capabilities. Just like most open-source releases, the main points in regards to the mannequin’s dataset is just not identified.