Xiaomi releases MiMo-v2.5 Family weights with strong coding and agent benchmarks

steveharing1

It was released around a week ago but Xiaomi Open Sourced it Now & its MIT Licensed and available om HuggingFace

landtuna

Ugh, the slop. "That's not a benchmark row. That's a story."

Alifatisk

I’ve been curious about MiMo models, they score so high on wide range of benchmarks, yet, not much is talked about it. At least in my bubble. Is it a serious contender? What is the model for?

swiftcoder

Has anyone run the numbers on Xiaomi's token subscription prices? The subscriptions are all priced in terms of "credits", and its not clear how credits map to tokens

esperent

I thought I had a high tolerance for slop but this is literally painful to read. Here's a slightly better source that looks like a human at least reviewed it before publishing. https://venturebeat.com/ai/open-source-xiaomi-mimo-v2-5-and-...

spwa4

TLDR: 2 models, 1 trillion parameters total, 42M active, 1M (MiMo-V2.5-Pro) or 256K (MiMo-V2.5-Pro-Base) context length, available on huggingface here: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro

steveharing1

Lets see when the community come up with more quantized versions. Waiting for Unsloth's version