It scored only 4% on the CritPt physics benchmark.
MiniMax-M3 is now the best open-source model on the AAI
It's ahead of Kimi-K2.6, GLM-5.1 and DeepSeek-V4-Pro
On the Coding Index it's slightly behind V4-Pro and K2.6, but it's beating them on the Agentic Index
CritPt scores are a bit worrisome.
But the biggest highlight here is probably its reasoning efficiency compared to other open models. (but still miles behind closed models)
Some users are surprised by MiniMax-M3 leading on the Artificial Analysis index and praise its agentic performance in real use, while many others dismiss the results as benchmaxxed, meaningless, or far behind closed models.
No Digg Deeper questions have been answered for this story yet.
Yeah it's probably a lot smaller than Kimi
MiniMax-M3 is now the best open-source model on the AAI
It's ahead of Kimi-K2.6, GLM-5.1 and DeepSeek-V4-Pro
On the Coding Index it's slightly behind V4-Pro and K2.6, but it's beating them on the Agentic Index
CritPt scores are a bit worrisome.
But the biggest highlight here is probably its reasoning efficiency compared to other open models. (but still miles behind closed models)
https://artificialanalysis.ai/models/minimax-m3
@scaling01 Open weight not open source 😶 And waiting to see the license
@scaling01 Don’t know what they did to V4-Pro but I don’t get the hype at all. Seems completely benchmaxxed spewing with overfitting from other benchmarks not including in AA like DeepSWE.
@scaling01 None of these match my experience. It thinks a lot about everything and the results are terrible and sloppy.
@scaling01 kimi still wins osworld and speed:
@scaling01 the coding index is just a lagging indicator what matters is the agentic score, where minimax-m3 is ahead of the pack. that's where the real action happens in ai development.
@scaling01 MiniMax M3 shipped coding benchmarks, multimodal, and 1M context together.
the M2 report already showed they test every attention variant and take the boring winner.
@scaling01 MiniMax is made to give high numbers on benchmarks, but it never translates into use
MiMo and Qwen for example, always feel better than their numbers
@scaling01 M2.7 → M3 is a solid jump ngl. but price-to-intelligence is what I care about and MiMo V2.5 Pro is still unbeaten.
@scaling01 I am curious how good models like this would be given more compute after seeing what Cursor turned the Kimi base into.
@scaling01 agentic index is cool but real reasoning efficiency battle is in sentient arena rn. challenge 0 got 5x more agent submissions in 2 weeks and it's all open source. gap closing faster than charts show
@scaling01 but still miles behind oai models (ftfy)
@scaling01 Is it open source ?
@scaling01 I’ve heard it’s benchmaxed
@scaling01 I didn't expect Minimax to straight out win against everyone
@scaling01 Its bullshit
@scaling01 So far until GLM5.5 comes along
@scaling01 It's benchmaxxed.
@scaling01 CritPt 那块还得再观望下