How much VRAM does llm-jp-4-32b-a3b-thinking need?

Q4_K_M is exactly 21,400,590,016 bytes (19.93 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is llm-jp-4-32b-a3b-thinking's KV cache?

2.00 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

Is llm-jp-4-32b-a3b-thinking a mixture-of-experts model?

Yes — 128 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.

llm-jp · text · mixture of experts

llm-jp-4-32b-a3b-thinking

llm-jp/llm-jp-4-32b-a3b-thinking

llm-jp-4-32b-a3b-thinking at Q4_K_M is exactly 21,400,590,016 bytes (19.93 GiB / 21.40 GB) — an effective 5.327 bits per weight, not the nominal 4. Its KV cache at 32K is 2.00 GiB.

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

Parameters

32.1B

total, not active

Architecture

qwen3moe

32 layers

Context

65,536

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q3_K_M	15.60 GiB	16,745,404,096	4.168	—	mmnga-o
IQ4_XS	16.37 GiB	17,582,299,072	4.377	—	hiratagoh
Q4_K_M	19.93 GiB	21,400,590,016	5.327	—	mmnga-o
Q4_K_M	19.93 GiB	21,400,590,272	5.327	—	hiratagoh
Q4_K_M	20.04 GiB	21,522,487,200	5.357	—	ash2813
Q5_K_M	22.73 GiB	24,407,381,696	6.075	—	mmnga-o
Q5_K_M	22.73 GiB	24,407,381,952	6.075	—	hiratagoh
Q6_K	26.86 GiB	28,840,728,512	7.179	—	hiratagoh
Q8_0	31.84 GiB	34,183,878,592	8.509	—	hiratagoh
BF16	59.89 GiB	64,304,223,936	16.006	—	hiratagoh
F163 shards	59.89 GiB	64,304,224,256	16.006	—	ash2813

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.25 GiB	0.25 GiB	—	32 / 0 / 0
8,192	0.50 GiB	0.50 GiB	—	32 / 0 / 0
16,384	1.00 GiB	1.00 GiB	—	32 / 0 / 0
32,768	2.00 GiB	2.00 GiB	—	32 / 0 / 0
65,536	4.00 GiB	4.00 GiB	—	32 / 0 / 0
131,072	8.00 GiB	8.00 GiB	—	32 / 0 / 0

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 16.84 GiB. The real file is 19.93 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

2560

Vocab

196,608

Sliding window

none

SWA period

—

MLA

Experts

128

Experts per token

use_sliding_window

false

Questions people ask

How much VRAM does llm-jp-4-32b-a3b-thinking need?: Q4_K_M is exactly 21,400,590,016 bytes (19.93 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is llm-jp-4-32b-a3b-thinking's KV cache?: 2.00 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 llm-jp-4-32b-a3b-thinking a mixture-of-experts model?: Yes — 128 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 llm-jp-4-32b-a3b-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.