How much VRAM does Mellum2-12B-A2.5B-Thinking need?

Q4_K_M is exactly 8,071,295,040 bytes (7.52 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Mellum2-12B-A2.5B-Thinking's KV cache?

0.50 GiB at 32K context with an f16 cache, computed per layer from the model's architecture — most of its layers use sliding-window attention, so a flat formula overstates this substantially. Quantizing the cache to q8_0 roughly halves it.

Is Mellum2-12B-A2.5B-Thinking a mixture-of-experts model?

Yes — 64 experts with 8 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

JetBrains · text · mixture of experts

Mellum2-12B-A2.5B-Thinking

JetBrains/Mellum2-12B-A2.5B-Thinking

Mellum2-12B-A2.5B-Thinking at Q4_K_M is exactly 8,071,295,040 bytes (7.52 GiB / 8.07 GB) — an effective 5.314 bits per weight, not the nominal 4. Its KV cache at 32K is 0.50 GiB, not the 1.75 GiB a flat formula predicts.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

12.1B

total, not active

Architecture

mellum

28 layers

Context

131,072

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q4_K_M	7.52 GiB	8,071,295,040	5.314	—	JetBrains
Q6_K	10.13 GiB	10,879,834,176	7.164	—	JetBrains
Q8_0	12.04 GiB	12,925,417,536	8.511	—	JetBrains

KV cache by context

computed per layer — this model uses sliding-window attention

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.12 GiB	0.22 GiB	1.88×	7 / 21 / 0
8,192	0.17 GiB	0.44 GiB	2.56×	7 / 21 / 0
16,384	0.28 GiB	0.88 GiB	3.12×	7 / 21 / 0
32,768	0.50 GiB	1.75 GiB	3.51×	7 / 21 / 0
65,536	0.94 GiB	3.50 GiB	3.74×	7 / 21 / 0
131,072	1.81 GiB	7.00 GiB	3.86×	7 / 21 / 0

21 of 28 layers cache only a 1,024-token window rather than the full context, on a period of 4. Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-12B-it-qat-q4_0-unquantized vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 6.36 GiB. The real file is 7.52 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 1.75 GiB at 32K context where the real figure is 0.50 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

2304

Vocab

98,304

Sliding window

1024

SWA period

MLA

Experts

Experts per token

use_sliding_window

true

Questions people ask

How much VRAM does Mellum2-12B-A2.5B-Thinking need?: Q4_K_M is exactly 8,071,295,040 bytes (7.52 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Mellum2-12B-A2.5B-Thinking's KV cache?: 0.50 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is Mellum2-12B-A2.5B-Thinking a mixture-of-experts model?: Yes — 64 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of Mellum2-12B-A2.5B-Thinking should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.