JetBrains’ open-source push this time is genuinely solid—its 12B model only activates 2.5B, with costs tightly controlled. The MTP draft acceleration is also kind of interesting.

JetBrains open-source Mellum-2, a 12B parameter code-mixed expert model. To control inference costs, only about 2.5 billion parameters are activated per token, with weights already released on Hugging Face, Apache 2.0. A new multi-token prediction (MTP) module has been added, where the MTP head acts as a draft model to accelerate sampling during inference. Provides three versions: basic, dialogue, and thinking; the thinking version can display explicit reasoning chains before output. Benchmarks: Humaneval 41.46%, MMLU 70.87%.