MiniMax releases M3, a 428B open-weight model with a 1M context window that scores 59% on SWE-Bench Pro

with only ~428B parameters and ~23B activated parameters Smaller than I expected

it's too small

I smell benchmaxxing

yep it's pretty small!

@MiniMax_AI @grok, does the 23Billion activated context mean this large model is theoretically usable on consumer hardware locally? GTX 4090 for example?

@MiniMax_AI @Sentdex they are here

@MiniMax_AI why isn't it benchmarked against the remaining open source guys like dsv4 pro, kimi k2.6/k2.7 code & so on

@zephyr_z9 only a little bit larger than DeepSeek v4 flash, 284B and 13B respectively.

@josepha_mayo @MiniMax_AI

@zephyr_z9 My stock

⏬Details.

@revolaition @MiniMax_AI loll

@MiniMax_AI 428B parameters 😭😭

@josepha_mayo @MiniMax_AI You can see it benchmarked agains kimi K2.6 on huggingface. Beats it on pretty much all at less than half size and moe https://huggingface.co/MiniMaxAI/MiniMax-M3

@zephyr_z9 My Latest Strategic Plan！..

👇

👇

@MiniMax_AI Congrats guys! Thank you MiniMax team for always releasing super smart open models! 🫡🥰

@SNARKAMOTO @MiniMax_AI Lmao

@teortaxesTex That’s 128 a100 40gb gpus for full context. I wouldn’t call this small

@MiniMax_AI Hoping for highspeed 🤞

@MiniMax_AI Congratulations! The best open model at half the size of the rest!

@MiniMax_AI This is perfect for 4x RTX 6000 Pros!