Mellum2 12B-A2.5B Thinking
JetBrains · Mellum · open weights
- GQA
- Sliding window
- RoPE
- Pre-norm
- QK-norm
- MoE
Facts and where they come from
| Released | 2026-05 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 12B | labmodel cardmodel name Mellum2-12B-A2.5B |
| Active parameters | 2.5B | labmodel cardmodel name ...-A2.5B |
| Context length | 128K tokens | labmodel cardREADME: Context Length: 131,072 |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for mellum: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for mellum: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | codemodelling codetransformers 5.18.0 mellum: q_norm and k_norm |
| Positional encoding | RoPE | codemodelling coderotary on the full head (default) |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for mellum: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
21× GQA 32q/4kv, window 1024 + 7× GQA 32q/4kv. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 12.1B |
|---|---|
| Active per token (modelled) | 2.44B |
| Without embeddings and output head | 11.7B total, 1.99B active |
| Published weights (Hugging Face count) | 12.1B |
| KV cache per token, BF16 (layers that grow with context) | 14 KiB |
| KV cache + state at 128K tokens, BF16 | 1.79 GiB |
| Decode FLOPs per token at 4K context | 5.25 GFLOP |
| Prefill FLOPs for a 4K prompt | 18.5 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 2,304 | config.jsonconfig.jsonhidden_size |
| vocab | 98,304 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 32 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 4 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.full.qk_norm | true | codemodelling codetransformers 5.18.0 mellum: q_norm and k_norm |
| mixers.sliding.type | attn | codemodelling codeattention block |
| mixers.sliding.heads | 32 | config.jsonconfig.jsonnum_attention_heads |
| mixers.sliding.kv_heads | 4 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.sliding.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.sliding.window | 1,024 | config.jsonconfig.jsonsliding_window |
| mixers.sliding.qk_norm | true | codemodelling codetransformers 5.18.0 mellum: q_norm and k_norm |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 7,168 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 64 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 896 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| layout | 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe · 3× sliding/moe · 1× full/moe | config.jsonconfig.jsonlayer_types, mlp_layer_types |
Sources
- config.json @ a731155
- model card
- arXiv 2605.31268
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “JetBrains Mellum2 Thinking (12B-A2.5B)” (name only; see about).