Minimax M3 - my first test results with it are *Fantastic*

This release feels especially important: Minimax M3 is now available on Openrouter https://openrouter.ai/minimax/minimax-m3 , as well as on https://www.minimax.io and Ollama. It's currently 1/2 price until June 7th, and this model is a significant improvement over previous versions.

The M3 version has a 1 million token context limit, and is multi-modal from the ground up.

All the other Minimax LLMs up to M2.7 have been open sourced, and the expectation is that M3 weights will be released soon.

BTW, https://www.minimax.io also offers a suite of other specialized types of models to generate music, video, voice, etc., as well as a coding harness made for their models.

Here are a 3D game and CRUD database example I built with Minimax M3 this morning:

https://com-pute.com/nick/space_invaders_3d_minimax3.html A 3D space invaders game which cost $.14 (a total of 14 cents, after 4 iterations) on Openrouter
http://1y1z.com:4274 *A Northwind CRUD database example which cost $.89 (89 cents), after 5 iterations to add lots of specific features such as datatable filtering, sorting, CSV and PDF downloads

I notice that Minimax M3 performs a lot more structured testing and automatically iterates based upon the results of those tests, and it provides explicitly verbose feedback about what it's doing during the agentic session (I used it in Pi).

Minimax is promoting the fact that their sparce attention routing improves long context performance with 9.7x faster prefill and 15.6x faster decoding at 1 million tokens.

I can't wait to try this locally!

I can confirm that MiniMax-M3-GGUF/UD-IQ4_NL does run on 2 clustered DGX Spark machines. I only loaded and tested it with a small 8k context limit, and haven't done any sort of optimization or software development with it yet, but did see that the 2 Asus GX10s were getting 9.5 tokens per second out of the gate with the IQ4_NL quant. With a 3rd DGX Spark added to the cluster, or using an IQ3_XXS quant on 2 clustered machines, this model should be very useable for challenging tasks, when needed.

One note: Minimax's sparse attention for 1 million token context is not yet included in the gguf files from Unsloth, so there's still some work that needs to be completed by the community before all features of the quantized open source MiniMax M3 version are completely ready - but this is a really promising initial test.