Runs on ≤12 GB · Hugging Face ↗
Not independently scored by Artificial Analysis — its base model JetBrains/Mellum2-12B-A2.5B-Base is the closest proxy
Released by JetBrains, Apache-2.0. The next version of Mellum 2 Thinking: same MellumForCausalLM architecture (12B MoE, 64 experts with 8 active, about 2.5B active parameters, 28 layers, sliding window on 3 of every 4 layers, 128K context). The new part is post-training, mostly reinforcement learning in real repositories with a shell and tools.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈1 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| MXFP4 | 7.0 GB |
| Q4_K_M | 8.1 GB |
| Q6_K | 10.9 GB |
| Q8_0 | 12.9 GB |
| BF16 | 24.3 GB |
JetBrains ships its own GGUF and stock llama.cpp loads the architecture, so no special build. About 8 GB at Q4_K_M: fits an 8-16 GB card. Read it next to the Mellum 2 12B card: same architecture, and the difference between the two is what the agentic RL added. It is a thinking model, so give it a token limit with room for the reasoning.