← sources

JetBrains Mellum 2.1 12B A2.5B MoE is a very good model.

source: reddit (reddit:LocalLLaMA)published 2026-10-14 15:47ingested 2026-10-15 02:42

retained text

It runs at around m4 16 at 60 tps at full 130k context, you can easily run a Q8 if you have a 16 GB GPU, I was able to get 140TPS on 9070xt, 

It’s an MoE so I think we can run it well on smaller cards too, 8GB, 12GB should be possible to get upwards of 30-40 tps. 

I have used it a bit, it can code, but I think having a bigger model plan tasks in detail would make it quite good for quick task execution.

document fetched 2026-10-15 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5uqcu/jetbrains_mellum_21_12b_a25b_moe_is_a_very_good