LLM-jp-4.1 32B-A3B Thinking VRAM requirements

LLM-jp-4.1 32B-A3B Thinking has 32.1B parameters, of which about 3.8B are used per token; all 128 experts still have to be in memory. With an 8K-token context and one request it needs about 21.0 GB of GPU memory at Q4_K_M, 34.0 GB at FP8 and 66.9 GB at FP16/BF16. The published weights take 59.9 GB (BF16). The smallest setup here that holds it at Q4_K_M with an 8K context is the RTX 3090 (24 GB), which runs it with up to 44K tokens of context.

21.0 GB at Q4_K_M, 8K context, one request

FP8
34.0 GB
FP16 / BF16
66.9 GB
Published weights
59.9 GB
Smallest setup, Q4_K_M
RTX 3090 (24 GB)
Open LLM-jp-4.1 32B-A3B Thinking in the calculator

Worked example: LLM-jp-4.1 32B-A3B Thinking with 32K tokens

Inputs: LLM-jp-4.1 32B-A3B Thinking · Q4_K_M weights · 32,768 tokens of context · one request · FP16 KV cache

Weights (32.1B at Q4_K_M)
18.1 GB
KV cache (64 KB per token)
2.00 GB
Buffers and runtime (0.5 GB + 10%)
2.51 GB
Total
22.6 GB

The smallest setup here that holds it is the RTX 3090 (24 GB), with about 1.38 GB to spare. Change the inputs in the calculator

What makes LLM-jp-4.1 32B-A3B Thinking's memory use different

As a mixture-of-experts model it reads about 2.16 GB of its 18.1 GB Q4_K_M weights per generated token (12%), so it writes like a much smaller model while needing memory for all of them.

How much VRAM does LLM-jp-4.1 32B-A3B Thinking need?

LLM-jp-4.1 32B-A3B Thinking needs 21.0 GB at Q4_K_M, 36.0 GB at Q8_0 and 66.9 GB at FP16 with 8,192 tokens of context and one request. Each total below is the weights plus the FP16 KV cache and the runtime overhead (0.5 GB plus 10%), and each row opens the calculator with that setting. What the GGUF names mean.

PrecisionWeightsTotalSmallest setup
As published (BF16) 59.9 GB 66.9 GB A100 80GB
FP16 / BF16 59.9 GB 66.9 GB A100 80GB
FP8 / INT8 29.9 GB 34.0 GB M4 Pro Mac (64 GB, 48 GB usable)
INT4 (AWQ / GPTQ) 15.9 GB 18.5 GB RTX 3090 (24 GB)
GGUF Q8_0 31.8 GB 36.0 GB M4 Pro Mac (64 GB, 48 GB usable)
GGUF Q6_K 24.5 GB 28.0 GB RTX 5090 (32 GB)
GGUF Q5_K_M 21.2 GB 24.4 GB RTX 5090 (32 GB)
GGUF Q4_K_M 18.1 GB 21.0 GB RTX 3090 (24 GB)
GGUF IQ4_XS 16.3 GB 19.0 GB RTX 3090 (24 GB)
GGUF Q3_K_M 14.6 GB 17.1 GB RTX 3090 (24 GB)
GGUF IQ3_XXS 12.3 GB 14.6 GB RTX 4060 Ti 16GB
GGUF Q2_K 12.5 GB 14.8 GB RTX 4060 Ti 16GB

Fine-tuning LLM-jp-4.1 32B-A3B Thinking? LLM-jp-4.1 32B-A3B Thinking VRAM for LoRA, QLoRA and full training.

LLM-jp-4.1 32B-A3B Thinking at 8K, 32K and 64K (full) tokens of context

All 32 layers use full attention. Each extra token of context adds 64 KB of FP16 cache per request. The last column is the smallest setup here, one card or a group, that holds it at Q4_K_M. How the KV cache works.

ContextKV cache, FP16Total at Q4_K_MTotal at Q8_0Smallest setup, Q4_K_M
8K tokens 512 MB 21.0 GB 36.0 GB RTX 3090 (24 GB)
32K tokens 2.00 GB 22.6 GB 37.7 GB RTX 3090 (24 GB)
64K tokens 4.00 GB 24.8 GB 39.9 GB RTX 5090 (32 GB)

How fast LLM-jp-4.1 32B-A3B Thinking writes

Tokens per second for one request with 8,192 tokens of context, estimated from memory bandwidth and the parameters read per token. A dash means it does not fit on one card. Try other settings in the speed calculator.

HardwareBandwidthQ4_K_MFP8
RTX 3060 12GB 360 GB/s ——
RTX 4090 1,008 GB/s 91–162—
RTX 5090 1,792 GB/s 147–271—
M4 Max Mac (128 GB) 546 GB/s 53–9136–61
M3 Ultra Mac Studio (512 GB) 819 GB/s 76–13452–90
H100 SXM 3,350 GB/s 231–454171–322
H200 4,800 GB/s 287–592221–431

Longest context on one GPU

How many tokens of context LLM-jp-4.1 32B-A3B Thinking fits on a single card with one request, an FP16 KV cache and 0.5 GB left free. "Full" means the model's whole context window fits.

GPUMemoryQ4_K_MQ8_0FP8
RTX 3060 12 GB NoNoNo
RTX 4060 Ti 16GB 16 GB NoNoNo
RTX 3090 / 4090 24 GB 44KNoNo
RTX 5090 32 GB 64K (full)NoNo
A100 40GB 40 GB 64K (full)58K64K (full)
Mac, 64 GB unified memory 48 GB 64K (full)64K (full)64K (full)
L40S / RTX 6000 Ada 48 GB 64K (full)64K (full)64K (full)
A100 / H100 80GB 80 GB 64K (full)64K (full)64K (full)
Mac, 128 GB unified memory 96 GB 64K (full)64K (full)64K (full)
H200 141 GB 64K (full)64K (full)64K (full)
B200 180 GB 64K (full)64K (full)64K (full)

Model details

Parameters
32.1B (32,139,028,992)
Experts
128 routed experts, all loaded
Active per token
3.8B
Layers
All 32 layers use full attention
Attention cache
4 KV heads × 128
Context length
65,536 tokens
Published weights
59.9 GB (BF16)
On Hugging Face
llm-jp/llm-jp-4.1-32b-a3b-thinking

Why this estimate looks this way

At Q4_K_M and an 8K-token context, LLM-jp-4.1 32B-A3B Thinking uses 18.1 GB for weights, 512 MB for its FP16 KV cache and 2.36 GB for estimated runtime overhead, totaling 21.0 GB. The overhead is a 0.5 GB base plus 10% of weights and cache.

This is a mixture-of-experts model: all 128 experts contribute to the 32.1B parameters held in memory, even though only about 3.8B parameters run per token. Using only active parameters would understate VRAM.

The architecture and published checkpoint size come from the model's config.json and weight files. These are estimates rather than measured peak memory; inference engines can reserve extra buffers or preallocate the full configured cache.

LLM-jp-4.1 32B-A3B Thinking next to similar-sized models

The three models here closest to it in parameter count, at Q4_K_M with 8,192 tokens of context.

ModelParametersTotalKV per tokenSmallest setup
LLM-jp-4.1 32B-A3B Thinking 32.1B, 3.8B active 21.0 GB 64 KB RTX 3090 (24 GB)
Nemotron 3 Nano 30B-A3B 31.6B, 3.5B active 20.1 GB 6 KB RTX 3090 (24 GB)
Gemma 4 31B 31.3B 21.9 GB 80 KB RTX 3090 (24 GB)
GLM-4.7 Flash 31.2B, 3B active 20.3 GB 53 KB RTX 3090 (24 GB)

Badge for your model card

Paste it into a Hugging Face or GitHub README; it shows the Q4_K_M total above and links to this page. Q8_0 and FP16 badges.

LLM-jp-4.1 32B-A3B Thinking VRAM: 21.0 GB at Q4_K_M, 8K context

Numbers read from the model files on Hugging Face on . Compare it with other models in the reproducible model and GPU report.