Dataset Viewer
Auto-converted to Parquet Duplicate
model_id
stringclasses
35 values
model
stringclasses
35 values
precision
stringclasses
9 values
file_label
stringclasses
13 values
context
int64
4.1k
131k
weights_gib
float64
1.81
1.91k
kv_cache_gib
float64
0.05
40
runtime_gib
float64
0.94
1.25
total_gib
float64
3.2
1.92k
smallest_single_card_gb
float64
8
288
⌀
qwen3.8-27b
Qwen3.8 27B
BF16
BF16
4,096
50.9
0.39
1.22
52.52
80
qwen3.8-27b
Qwen3.8 27B
BF16
BF16
8,192
50.9
0.64
1.22
52.77
80
qwen3.8-27b
Qwen3.8 27B
BF16
BF16
32,768
50.9
2.14
1.22
54.27
80
qwen3.8-27b
Qwen3.8 27B
BF16
BF16
131,072
50.9
8.14
1.22
60.27
80
qwen3.8-27b
Qwen3.8 27B
FP8
FP8
4,096
28.75
0.39
1.22
30.36
32
qwen3.8-27b
Qwen3.8 27B
FP8
FP8
8,192
28.75
0.64
1.22
30.61
32
qwen3.8-27b
Qwen3.8 27B
FP8
FP8
32,768
28.75
2.14
1.22
32.11
48
qwen3.8-27b
Qwen3.8 27B
FP8
FP8
131,072
28.75
8.14
1.22
38.11
48
qwen3.8-27b
Qwen3.8 27B
Q8_0
Q8_0
4,096
27.05
0.39
1.22
28.67
32
qwen3.8-27b
Qwen3.8 27B
Q8_0
Q8_0
8,192
27.05
0.64
1.22
28.92
32
qwen3.8-27b
Qwen3.8 27B
Q8_0
Q8_0
32,768
27.05
2.14
1.22
30.42
32
qwen3.8-27b
Qwen3.8 27B
Q8_0
Q8_0
131,072
27.05
8.14
1.22
36.42
48
qwen3.8-27b
Qwen3.8 27B
Q6_K
Q6_K
4,096
20.47
0.39
1.22
22.09
24
qwen3.8-27b
Qwen3.8 27B
Q6_K
Q6_K
8,192
20.47
0.64
1.22
22.34
24
qwen3.8-27b
Qwen3.8 27B
Q6_K
Q6_K
32,768
20.47
2.14
1.22
23.84
32
qwen3.8-27b
Qwen3.8 27B
Q6_K
Q6_K
131,072
20.47
8.14
1.22
29.84
32
qwen3.8-27b
Qwen3.8 27B
Q5_K_M
Q5_K_M
4,096
18.41
0.39
1.22
20.03
24
qwen3.8-27b
Qwen3.8 27B
Q5_K_M
Q5_K_M
8,192
18.41
0.64
1.22
20.28
24
qwen3.8-27b
Qwen3.8 27B
Q5_K_M
Q5_K_M
32,768
18.41
2.14
1.22
21.78
24
qwen3.8-27b
Qwen3.8 27B
Q5_K_M
Q5_K_M
131,072
18.41
8.14
1.22
27.78
32
qwen3.8-27b
Qwen3.8 27B
Q4_K_M
Q4_K_M
4,096
15.33
0.39
1.22
16.95
24
qwen3.8-27b
Qwen3.8 27B
Q4_K_M
Q4_K_M
8,192
15.33
0.64
1.22
17.2
24
qwen3.8-27b
Qwen3.8 27B
Q4_K_M
Q4_K_M
32,768
15.33
2.14
1.22
18.7
24
qwen3.8-27b
Qwen3.8 27B
Q4_K_M
Q4_K_M
131,072
15.33
8.14
1.22
24.7
32
qwen3.8-27b
Qwen3.8 27B
Q2_K
UD-Q2_K_XL
4,096
9.15
0.39
1.22
10.77
12
qwen3.8-27b
Qwen3.8 27B
Q2_K
UD-Q2_K_XL
8,192
9.15
0.64
1.22
11.02
12
qwen3.8-27b
Qwen3.8 27B
Q2_K
UD-Q2_K_XL
32,768
9.15
2.14
1.22
12.52
16
qwen3.8-27b
Qwen3.8 27B
Q2_K
UD-Q2_K_XL
131,072
9.15
8.14
1.22
18.52
24
qwen3.5-9b
Qwen3.5 9B
BF16
BF16
4,096
16.69
0.17
1.22
18.09
24
qwen3.5-9b
Qwen3.5 9B
BF16
BF16
8,192
16.69
0.3
1.22
18.21
24
qwen3.5-9b
Qwen3.5 9B
BF16
BF16
32,768
16.69
1.05
1.22
18.96
24
qwen3.5-9b
Qwen3.5 9B
BF16
BF16
131,072
16.69
4.05
1.22
21.96
24
qwen3.5-9b
Qwen3.5 9B
Q8_0
Q8_0
4,096
8.87
0.17
1.22
10.27
12
qwen3.5-9b
Qwen3.5 9B
Q8_0
Q8_0
8,192
8.87
0.3
1.22
10.39
12
qwen3.5-9b
Qwen3.5 9B
Q8_0
Q8_0
32,768
8.87
1.05
1.22
11.14
12
qwen3.5-9b
Qwen3.5 9B
Q8_0
Q8_0
131,072
8.87
4.05
1.22
14.14
16
qwen3.5-9b
Qwen3.5 9B
Q6_K
Q6_K
4,096
6.95
0.17
1.22
8.34
12
qwen3.5-9b
Qwen3.5 9B
Q6_K
Q6_K
8,192
6.95
0.3
1.22
8.47
12
qwen3.5-9b
Qwen3.5 9B
Q6_K
Q6_K
32,768
6.95
1.05
1.22
9.22
12
qwen3.5-9b
Qwen3.5 9B
Q6_K
Q6_K
131,072
6.95
4.05
1.22
12.22
16
qwen3.5-9b
Qwen3.5 9B
Q5_K_M
Q5_K_M
4,096
6.13
0.17
1.22
7.52
8
qwen3.5-9b
Qwen3.5 9B
Q5_K_M
Q5_K_M
8,192
6.13
0.3
1.22
7.65
8
qwen3.5-9b
Qwen3.5 9B
Q5_K_M
Q5_K_M
32,768
6.13
1.05
1.22
8.4
12
qwen3.5-9b
Qwen3.5 9B
Q5_K_M
Q5_K_M
131,072
6.13
4.05
1.22
11.4
12
qwen3.5-9b
Qwen3.5 9B
Q4_K_M
Q4_K_M
4,096
5.29
0.17
1.22
6.69
8
qwen3.5-9b
Qwen3.5 9B
Q4_K_M
Q4_K_M
8,192
5.29
0.3
1.22
6.81
8
qwen3.5-9b
Qwen3.5 9B
Q4_K_M
Q4_K_M
32,768
5.29
1.05
1.22
7.56
8
qwen3.5-9b
Qwen3.5 9B
Q4_K_M
Q4_K_M
131,072
5.29
4.05
1.22
10.56
12
qwen3.5-9b
Qwen3.5 9B
Q3_K_M
Q3_K_M
4,096
4.35
0.17
1.22
5.75
8
qwen3.5-9b
Qwen3.5 9B
Q3_K_M
Q3_K_M
8,192
4.35
0.3
1.22
5.87
8
qwen3.5-9b
Qwen3.5 9B
Q3_K_M
Q3_K_M
32,768
4.35
1.05
1.22
6.62
8
qwen3.5-9b
Qwen3.5 9B
Q3_K_M
Q3_K_M
131,072
4.35
4.05
1.22
9.62
12
qwen3.5-9b
Qwen3.5 9B
Q2_K
UD-Q2_K_XL
4,096
3.84
0.17
1.22
5.23
8
qwen3.5-9b
Qwen3.5 9B
Q2_K
UD-Q2_K_XL
8,192
3.84
0.3
1.22
5.36
8
qwen3.5-9b
Qwen3.5 9B
Q2_K
UD-Q2_K_XL
32,768
3.84
1.05
1.22
6.11
8
qwen3.5-9b
Qwen3.5 9B
Q2_K
UD-Q2_K_XL
131,072
3.84
4.05
1.22
9.11
12
qwen3.5-4b
Qwen3.5 4B
BF16
BF16
4,096
7.85
0.17
1.22
9.24
12
qwen3.5-4b
Qwen3.5 4B
BF16
BF16
8,192
7.85
0.3
1.22
9.37
12
qwen3.5-4b
Qwen3.5 4B
BF16
BF16
32,768
7.85
1.05
1.22
10.12
12
qwen3.5-4b
Qwen3.5 4B
BF16
BF16
131,072
7.85
4.05
1.22
13.12
16
qwen3.5-4b
Qwen3.5 4B
Q8_0
Q8_0
4,096
4.17
0.17
1.22
5.57
8
qwen3.5-4b
Qwen3.5 4B
Q8_0
Q8_0
8,192
4.17
0.3
1.22
5.7
8
qwen3.5-4b
Qwen3.5 4B
Q8_0
Q8_0
32,768
4.17
1.05
1.22
6.45
8
qwen3.5-4b
Qwen3.5 4B
Q8_0
Q8_0
131,072
4.17
4.05
1.22
9.45
12
qwen3.5-4b
Qwen3.5 4B
Q6_K
Q6_K
4,096
3.28
0.17
1.22
4.68
8
qwen3.5-4b
Qwen3.5 4B
Q6_K
Q6_K
8,192
3.28
0.3
1.22
4.8
8
qwen3.5-4b
Qwen3.5 4B
Q6_K
Q6_K
32,768
3.28
1.05
1.22
5.55
8
qwen3.5-4b
Qwen3.5 4B
Q6_K
Q6_K
131,072
3.28
4.05
1.22
8.55
12
qwen3.5-4b
Qwen3.5 4B
Q5_K_M
Q5_K_M
4,096
2.93
0.17
1.22
4.32
8
qwen3.5-4b
Qwen3.5 4B
Q5_K_M
Q5_K_M
8,192
2.93
0.3
1.22
4.45
8
qwen3.5-4b
Qwen3.5 4B
Q5_K_M
Q5_K_M
32,768
2.93
1.05
1.22
5.2
8
qwen3.5-4b
Qwen3.5 4B
Q5_K_M
Q5_K_M
131,072
2.93
4.05
1.22
8.2
12
qwen3.5-4b
Qwen3.5 4B
Q4_K_M
Q4_K_M
4,096
2.55
0.17
1.22
3.95
8
qwen3.5-4b
Qwen3.5 4B
Q4_K_M
Q4_K_M
8,192
2.55
0.3
1.22
4.07
8
qwen3.5-4b
Qwen3.5 4B
Q4_K_M
Q4_K_M
32,768
2.55
1.05
1.22
4.82
8
qwen3.5-4b
Qwen3.5 4B
Q4_K_M
Q4_K_M
131,072
2.55
4.05
1.22
7.82
12
qwen3.5-4b
Qwen3.5 4B
Q3_K_M
Q3_K_M
4,096
2.14
0.17
1.22
3.53
8
qwen3.5-4b
Qwen3.5 4B
Q3_K_M
Q3_K_M
8,192
2.14
0.3
1.22
3.66
8
qwen3.5-4b
Qwen3.5 4B
Q3_K_M
Q3_K_M
32,768
2.14
1.05
1.22
4.41
8
qwen3.5-4b
Qwen3.5 4B
Q3_K_M
Q3_K_M
131,072
2.14
4.05
1.22
7.41
8
qwen3.5-4b
Qwen3.5 4B
Q2_K
UD-Q2_K_XL
4,096
1.81
0.17
1.22
3.2
8
qwen3.5-4b
Qwen3.5 4B
Q2_K
UD-Q2_K_XL
8,192
1.81
0.3
1.22
3.33
8
qwen3.5-4b
Qwen3.5 4B
Q2_K
UD-Q2_K_XL
32,768
1.81
1.05
1.22
4.08
8
qwen3.5-4b
Qwen3.5 4B
Q2_K
UD-Q2_K_XL
131,072
1.81
4.05
1.22
7.08
8
qwen3-8b
Qwen3 8B
BF16
BF16
4,096
15.26
0.56
1.04
16.86
24
qwen3-8b
Qwen3 8B
BF16
BF16
8,192
15.26
1.13
1.04
17.43
24
qwen3-8b
Qwen3 8B
BF16
BF16
32,768
15.26
4.5
1.04
20.8
24
qwen3-8b
Qwen3 8B
FP8
FP8
4,096
8.79
0.56
1.04
10.39
12
qwen3-8b
Qwen3 8B
FP8
FP8
8,192
8.79
1.13
1.04
10.95
12
qwen3-8b
Qwen3 8B
FP8
FP8
32,768
8.79
4.5
1.04
14.33
16
qwen3-8b
Qwen3 8B
Q8_0
Q8_0
4,096
8.11
0.56
1.04
9.71
12
qwen3-8b
Qwen3 8B
Q8_0
Q8_0
8,192
8.11
1.13
1.04
10.28
12
qwen3-8b
Qwen3 8B
Q8_0
Q8_0
32,768
8.11
4.5
1.04
13.65
16
qwen3-8b
Qwen3 8B
Q6_K
Q6_K
4,096
6.26
0.56
1.04
7.87
12
qwen3-8b
Qwen3 8B
Q6_K
Q6_K
8,192
6.26
1.13
1.04
8.43
12
qwen3-8b
Qwen3 8B
Q6_K
Q6_K
32,768
6.26
4.5
1.04
11.8
16
qwen3-8b
Qwen3 8B
Q5_K_M
Q5_K_M
4,096
5.45
0.56
1.04
7.05
8
qwen3-8b
Qwen3 8B
Q5_K_M
Q5_K_M
8,192
5.45
1.13
1.04
7.61
8
qwen3-8b
Qwen3 8B
Q5_K_M
Q5_K_M
32,768
5.45
4.5
1.04
10.99
12
qwen3-8b
Qwen3 8B
Q4_K_M
Q4_K_M
4,096
4.68
0.56
1.04
6.28
8
End of preview. Expand in Data Studio

LLM VRAM requirements: real GGUF sizes for 35 open models

How much GPU memory 35 current open LLMs need to run one request, at every published quantisation and at 4k, 8k, 32k and 128k context. The weights are the real sizes of the published files, not parameters × bits ÷ 8; the KV cache is computed from each model's own config.json, including hybrid, sliding-window and MLA attention, where the naive formula is wrong.

Snapshot: October 2026. The same data drives the VRAM calculator and the article LLM VRAM requirements: table of 35 models, which explains the findings.

Try it interactively in the Space gpunz/vram-calculator: pick a model, quantisation, context and parallel requests, and see which GPU it fits.

Files

llm_vram.csv: one row per model × precision × context (882 rows).

Column Meaning
model_id, model Model id and name
precision BF16, FP8, Q8_0 … Q2_K, or native for a release shipped already quantised (GPT-OSS in MXFP4)
file_label The file the size comes from, e.g. UD-Q4_K_XL where unsloth's dynamic quant stands in for Q4_K_M
context Context length, tokens
weights_gib Size of the weight files, GiB
kv_cache_gib KV cache (plus linear-attention state) for one sequence at FP16, GiB
runtime_gib CUDA context and buffers: about 0.75 GiB plus the logits of a 512-token batch
total_gib Sum of the three
smallest_single_card_gb Smallest common card (8, 12, 16, 24, 32, 48, 80, 96, 141, 180, 288 GB) where the total fits in 97.5% of memory; empty when no single card does

models.csv: one row per model: original Hugging Face repo, total and active parameters, maximum context, vocabulary, attention type and the per-token KV-cache sizes used for the estimate.

How the numbers are computed

weights   = size of the published file for that quantisation
kv_cache  = tokens × full_attention_bytes_per_token
          + min(tokens, window) × sliding_bytes_per_token
          + linear_attention_state
runtime   ≈ 0.75 GiB + 512 × vocab × 4 bytes
total     = weights + kv_cache + runtime
  • GGUF sizes come from builds by unsloth, ggml-org, bartowski and the model authors, read from the Hugging Face API. Rows whose size would have to be estimated from the parameter count are left out.
  • All sizes are GiB (1024³ bytes), the unit GPU memory is sold in, so they compare directly with "24 GB". Hugging Face lists decimal gigabytes, which is why its file sizes look about 7% larger.
  • One request, FP16 KV cache. A quantised KV cache (q8_0) halves the cache; parallel requests multiply it.
  • Real-world use adds a little: vLLM pre-allocates most of the card for its cache by design, so nvidia-smi always looks full there. These numbers are what the model needs.

Example

from datasets import load_dataset

df = load_dataset("gpunz/llm-vram-requirements", "llm_vram", split="train").to_pandas()
df[(df.precision == "Q4_K_M") & (df.context == 32768) & (df.total_gib <= 23.4)][["model", "total_gib"]]

License and source

CC BY 4.0. Compiled by gpu.nz, a GPU cloud billed by the minute. Corrections are welcome in the Community tab.

Downloads last month
67

Space using gpunz/llm-vram-requirements 1