I think the sentence “Larger than M2 (200B model)” is simply pointing out that M2.x was around a 200B parameter model. The new model is multimodal, so I’d expect it to be at least twice the size of M2, but definitely under 1T parameters, since training a trillion-parameter model would be extremely costly.
Step-3.7-Flash went multimodal compared to 3.5 with a simple 3.5GB mmproj.
It’s bigger but multimodality doesn’t require explosion in size.
Lets hope it can comfortably fit on dual sparks. In my limited testing on the minimax agent website the model feels better than m2.7 but doesn’t feel 500b class with its intelligence.
Supposed to be out tomorrow some time. But not VLLM PR so they could delay the model again like they did with M 2.7. Will see.
They already published some code GitHub - MiniMax-AI/MSA · GitHub
Really hope the license is good and not so vauge like the last one that making a single dollar off a output from M 3 from your small business falls under the no commercial use term. Most people running a model on $3k-$15k hardware are trying to get some benefit out of the models they run and not just “make me a GTA clone with a synthwave style” type prompts.
I’m sure it isn’t malicious, just avoiding scope creep so they can efficiently get day 0 support.
As soon as that PR is ported to vLLM we can just pull it in.
MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
850gb so we’ll have to wait for quants, most likely won’t fit 2x cluster but plenty of room in a 4x
Chunky, 428B-A23B - I was wrong on the size, that’s going to be difficult to fit on 2x Sparks.
I knew it! This confirms my initial expectations: the footprint is double that of the MiniMax 2.7 model. I am currently evaluating whether it can be successfully deployed across my two Sparks
It seems that the Qwen 397B, which is close in scale but slightly smaller, barely fits into a dual-Spark setup using intel INT4 quantization as it is. MiniMax is a bit larger, so it feels like it won’t fit into two Sparks at all.
maybe REAP version…
Probably won’t fit 2x DGX Sparks. Perhaps NVFP4. But probably 3 should be the good fit.
Let’s wait and see