Xiaomi has formally entered the unreal intelligence area with the discharge of its first open-source reasoning mannequin, MiMo 7B. Developed by the corporate’s newly established Huge Mannequin Core Group, MiMo represents a strategic shift from Xiaomi’s conventional {hardware} focus to superior AI analysis and improvement.
Regardless of having simply 7 billion parameters, MiMo has demonstrated sturdy efficiency in complicated reasoning duties, surpassing a lot bigger fashions like OpenAI’s o1-mini and Alibaba’s 32B QwQ-Preview. This efficiency has positioned Xiaomi within the highlight for producing a extremely environment friendly and compact mannequin able to rivaling extra resource-intensive opponents.
Benchmark Efficiency and Design Technique
MiMo 7B has outperformed its friends on benchmarks comparable to AIME 24-25 (a check for mathematical reasoning) and LiveCodeBench v5 (a programming problem dataset). This success is attributed to a well-structured improvement course of consisting of each pre-training and post-training improvements.
Pre-training Methods Included:
- Wealthy Reasoning Corpus: Targeted on extracting and integrating complicated reasoning data.
- Artificial Knowledge Era: Produced roughly 200 billion tokens of expert-level reasoning knowledge to enhance coaching depth.
- Progressive Problem Coaching: Employed a three-phase coaching methodology with rising ranges of issue.
- Intensive Token Publicity: Skilled throughout an enormous 25 trillion tokens, making certain complete studying.
Put up-training Enhancements:
- Take a look at Problem-Pushed Rewards: Launched a novel strategy for addressing reward sparsity in complicated algorithmic duties.
- Knowledge Re-sampling Methods: Utilized to stabilize reinforcement studying processes.
- Seamless Rollout System: Elevated coaching effectivity by 2.29 occasions and validation pace by 1.96 occasions, streamlining the reinforcement studying (RL) pipeline.
A Broader Imaginative and prescient for AI
The discharge of MiMo underscores Xiaomi’s broader ambitions within the AI sector. Whereas beforehand identified for shopper electronics and good units, Xiaomi is now positioning itself as a critical participant in AI mannequin improvement. By open-sourcing MiMo, Xiaomi is contributing to the collaborative AI analysis ecosystem, fostering innovation past proprietary improvement.
Builders and AI researchers can now entry MiMo 7B and its full technical documentation through Xiaomi’s official Hugging Face repository, providing a useful software for additional experimentation and improvement. This transfer displays Xiaomi’s intent to construct a robust presence in AI whereas supporting the open-source neighborhood.
Filed in . Learn extra about AI (Synthetic Intelligence), ChatGPT and Xiaomi.
