🤖MiniMax M3 Open Model Beats GPT-5.5 at 5-10% of the Cost
TL;DR
China's MiniMax released M3, an open-weight model that tops GPT-5.5 and Gemini 3.1 Pro on key benchmarks while costing 5-10% as much to run. It pairs a 1M-token context with sparse attention that cuts per-token compute to a twentieth of the prior generation.
China's MiniMax released M3, an open-weight model that tops GPT-5.5 and Gemini 3.1 Pro on key benchmarks while costing 5-10% as much to run. It pairs a 1M-token context with sparse attention that cuts per-token compute to a twentieth of the prior generation.
Key Points
Scores 59.0% on open-weight SWE-Bench Pro, rivaling proprietary leaders
1M-token context; sparse attention cuts per-token compute about 20x
Roughly 9x faster prefill and 15x faster decode vs the prior MiniMax model
Natively multimodal; pretraining corpus exceeds 100 trillion tokens
Open weights, aimed at making long-context agents economical to deploy
Why It Matters
Cheap, open, long-context models out of China keep dropping the price floor that US labs have to price against.
Quick Facts
Frequently Asked Questions
Why does this matter?
Cheap, open, long-context models out of China keep dropping the price floor that US labs have to price against.
What happened?
China's MiniMax released M3, an open-weight model that tops GPT-5.5 and Gemini 3.1 Pro on key benchmarks while costing 5-10% as much to run. It pairs a 1M-token context with sparse attention that cuts per-token compute to a twentieth of the prior generation.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 1,970 builders reading daily.