VL
VECTOR LAB

EST. 2025

WEEKLY UPDATE 2026
BY ANDREW MEAD

Minimax M3

MiniMax M3 release and the new best open source image gen model

Read in
Minimax M3 - Vector Lab
WEEKLY UPDATE 2026
BY ANDREW MEAD

Minimax M3

MiniMax M3 release and the new best open source image gen model

Read in

Releases

MiniMax M3

MiniMax is back with a new major model release. Their Minimax M3 model is their new flagship model meant to replace the fast and cheap MiniMax M2.7.

Benchmarks for Minimax

The model has not been open sourced yet, but will be sometime next week. Even though we don’t have exact numbers on the size of it, we do know that it is bigger given that the API price for it has doubled.

It does not stand out capability wise from any of the other Chinese options available right now, especially given the price, which used to at least differentiate it from GLM and Kimi. Now DeepSeek is the only truly cheap option, while the other labs all seem to be congregating around the $3-$4 price range.

I have seen many reports of users complaining baout how slow the model is. From the Artificial Analysis benchmarks it does not seem to over think, at least not more than any other Chinese model, which points to an API level issue, where the MiniMax team is struggling to serve the model at scale.

This is further corroborated by my new favorite SWE benchmark DeepSWE, where they said they were encountering issues running the model on their benchmark which is why we don’t have a score for it yet.

Model$ per million (input)$ per million (output)Tokens per second
GPT 5.4 Mini$0.75$4.562
Claude Sonnet 4.6$3$1537
Gemini 3.5 Flash$1.50$980
DeepSeek V4$0.43$0.8741
GLM 5.1$1.40$4.4033
MiniMax M2.7$0.30$1.2034
MiniMax M3$1.20$2.4040
Costs and speeds gotten from OpenRouter

If you are MiniMax/ Chinese model enjoyer I would give this model a look, but also be aware that it probably won’t blow you away with its performance versus any of the other Chinese models.

Quick Hits

Ideogram 4.0

There’s a new open source image gen model that is able to compete with the likes of Nano Banana and GPT Image. Ideogram 4.0 is from an independent lab, and is a 9.3 billion parameter diffusion model.

design arena eval

Design arena scores

This is by far the best local image gen model, taking the throne from the (relatively) outdated Flux 2 models. From what I have seen it has exceptional text rendering and aesthetic capabilities.

A bird in the hand


text rendering


The fight


Dog wearing headphones

Example images from their announcement post

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

Superhuman air hockey AI robot from Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS