MiMo-Audio claims few-shot audio generalization from 100M hours of pretraining
- Source
- x.com
- Date
👋 Say Hi to MiMo-Audio! Our BREAKTHROUGH in general-purpose audio intelligence. 🎯 Scaling pretraining to 100M+ hours leads to EMERGENCE of few-shot generalization across diverse audio tasks! 🔥 Post-trained MiMo-Audio-7B-Instruct: • crushes benchmarks: SOTA on MMSU, MMAU, MMAR, MMAU-Pro • outperforms Gemini-2.5-Flash on audio understanding • beats GPT-4o-Audio on complex reasoning tasks 💎 The best part? It's 100% OPEN-SOURCE Everything from tokenizer to model to evaluations! 🤗 Try it in HF Space: https://t.co/9EwFiHFdkt 📝 Tech Blog:
Context
Xiaomi's MiMo lab describes MiMo-Audio as a 7-billion-parameter audio- trained on more than 100 million hours of speech, a scale the team's technical report says produces few-shot generalization across audio tasks the model wasn't explicitly trained on. Xiaomi released the tokenizer, model checkpoints and its suite openly rather than keeping them behind an API.
The tweet's claims that the instruct model beats Gemini 2.5 Flash and GPT-4o-Audio on specific benchmarks are Xiaomi's own reported comparisons; no independent leaderboard result is cited in the source to confirm that ranking.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
Checking sign-in…
Loading comments…