Active Parameters
12B
Context Length
131K
Modality
Text
Architecture
Mixture of Experts
KV Cache Rate
56 KB/tok
License
Apache 2.0
Release Date
20 Sept 2026
Knowledge Cutoff
-
API Pricing (per 1M)
Self-hosted only
Precision
Context
Workload Memory: 9.69 GB
INT4 · 8K context
Base Weights: 6.75 GB
KV Cache: 0.44 GB
Runtime & Graphs: 2.50 GB
1x NVIDIA L4 24GB
9.7 / 21.6 GB usable
1x RTX 3060 12GB
9.7 / 10.8 GB usable
Apple M4 24GB
9.7 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Checkpoint format
FP16
CUDA
>= 11.8
PyTorch
>= 2.1
Transformers
>= 5.8.1
| Architecture | Precision | Attention kernel |
|---|---|---|
Blackwell, Hopper | Native FP16 | FlashAttention-3 |
Ada Lovelace, Ampere | Native FP16 | FlashAttention-2 |
No evaluation benchmarks for Mellum 2.1 12B A2.5B Thinking available.
Overall Rank
-
Coding Rank
-
Mellum 2.1 12B A2.5B Thinking is an MoE reasoning and coding model developed by JetBrains with 2.5B active parameters per token. It is tailored for advanced reasoning traces and software engineering workflows in IDE environments.
Attention
Attention Structure
Grouped-Query Attention
Attention Heads
32
Key-Value Heads
4
Attention Head Dimension
128
Position Embedding
ROPE
RoPE Theta
500,000
Sliding Window Attention
Yes
Sliding Window Size
1,024
Sliding Window Ratio
75.0%
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Base Precision
FP16
Hidden Dimension Size
2,304
Number of Layers
28
FFN Intermediate Size (Dense)
7,168
Tokenizer
Vocabulary Size
98,304
Mixture of Experts
Total Expert Parameters
2.5B
Number of Experts
64
Active Experts
8
Shared Experts
0
FFN Intermediate Size (per Expert)
896
Dense Layers Before MoE
0
The Mellum model family developed by JetBrains.