Dataset Viewer
Auto-converted to Parquet Duplicate
model_slug
stringlengths
7
20
model_name
stringlengths
3
23
creator
stringlengths
3
12
parameters
stringlengths
2
9
quantization
stringclasses
4 values
gguf_repo
stringlengths
18
48
source_repo
stringlengths
11
45
file_gb
float64
0.6
1.22k
max_context_tokens
float64
41k
1.05M
⌀
attention_shape_known
stringclasses
2 values
context_cache_gb_8k
float64
0.09
2.5
⌀
runtime_gb_8k
float64
0.9
37.5
⌀
total_gb_8k
float64
1.7
1.26k
⌀
context_cache_gb_32k
float64
0.34
10
⌀
runtime_gb_32k
float64
1.6
38.2
⌀
total_gb_32k
float64
2.7
1.26k
⌀
context_cache_gb_64k
float64
0.69
20
⌀
runtime_gb_64k
float64
2.6
39.2
⌀
total_gb_64k
float64
4.1
1.27k
⌀
context_cache_gb_128k
float64
1.38
40
⌀
runtime_gb_128k
float64
4.6
41.2
⌀
total_gb_128k
float64
6.8
1.27k
⌀
read_on
stringdate
2026-11-12 00:00:00
2026-11-12 00:00:00
deepseek-r1
DeepSeek-R1
DeepSeek
685B MoE
Q4_K_M
unsloth/DeepSeek-R1-GGUF
deepseek-ai/DeepSeek-R1
376.7
163,840
yes
0.54
12.2
389.4
2.14
12.9
391.7
4.29
13.9
394.9
8.58
15.9
401.2
2026-11-12
deepseek-v4-flash
DeepSeek-V4-Flash
DeepSeek
304B MoE
Q4_K
unsloth/DeepSeek-V4-Flash-0731-GGUF
deepseek-ai/DeepSeek-V4-Flash-0731
144.4
1,048,576
yes
0.38
5.2
150
1.51
5.9
151.8
3.02
6.9
154.4
6.05
8.9
159.4
2026-11-12
deepseek-v4-pro
DeepSeek-V4-Pro
DeepSeek
1650B MoE
Q4_K
unsloth/DeepSeek-V4-Pro-0813-GGUF
deepseek-ai/DeepSeek-V4-Pro-0813
791.3
1,048,576
yes
0.54
24.6
816.4
2.14
25.3
818.8
4.29
26.3
821.9
8.58
28.3
828.2
2026-11-12
devstral-small-2-24b
Devstral Small 2 24B
Mistral AI
24B
Q4_K_M
unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF
mistralai/Devstral-Small-2-24B-Instruct-2512
13.3
393,216
yes
1.25
1.2
15.8
5
2
20.3
10
3
26.3
20
5
38.3
2026-11-12
gemma-3-27b
Gemma 3 27B
Google
27.4B
Q4_K_M
unsloth/gemma-3-27b-it-GGUF
google/gemma-3-27b-it
15.4
131,072
yes
1.03
1.3
17.7
2.91
2.1
20.4
5.41
3.1
23.9
10.41
5.1
30.9
2026-11-12
gemma-4-12b
Gemma 4 12B
Google
12B
Q4_K_M
unsloth/gemma-4-12B-it-GGUF
google/gemma-4-12B-it
6.6
262,144
yes
0.44
1
8.1
0.81
1.8
9.2
1.31
2.8
10.7
2.31
4.8
13.7
2026-11-12
gemma-4-26b-a4b
Gemma 4 26B-A4B
Google
25.8B MoE
Q4_K_M
unsloth/gemma-4-26B-A4B-it-GGUF
google/gemma-4-26B-A4B-it
15.8
262,144
yes
0.35
1.3
17.5
0.82
2.1
18.7
1.45
3.1
20.3
2.7
5.1
23.6
2026-11-12
gemma-4-31b
Gemma 4 31B
Google
31.3B
Q4_K_M
unsloth/gemma-4-31B-it-GGUF
google/gemma-4-31B-it
17.1
262,144
yes
1.41
1.4
19.9
3.28
2.1
22.5
5.78
3.1
26
10.78
5.1
33
2026-11-12
gemma-4-e4b
Gemma 4 E4B
Google
8B
Q4_K_M
unsloth/gemma-4-E4B-it-GGUF
google/gemma-4-E4B-it
4.6
131,072
yes
0.25
1
5.8
0.91
1.7
7.2
1.78
2.7
9.1
3.53
4.7
12.9
2026-11-12
glm-4-5-air
GLM-4.5-Air
Zhipu AI
110B MoE
Q4_K_M
unsloth/GLM-4.5-Air-GGUF
zai-org/GLM-4.5-Air
68
131,072
yes
1.44
2.9
72.3
5.75
3.6
77.4
11.5
4.6
84.1
23
6.6
97.6
2026-11-12
glm-4-7-flash
GLM-4.7-Flash
Zhipu AI
31.2B MoE
Q4_K_M
unsloth/GLM-4.7-Flash-GGUF
zai-org/GLM-4.7-Flash
17.1
202,752
yes
0.41
1.4
18.9
1.65
2.1
20.9
3.3
3.1
23.5
6.61
5.1
28.8
2026-11-12
glm-5-2
GLM-5.2
Zhipu AI
753B MoE
Q4_K_M
unsloth/GLM-5.2-GGUF
zai-org/GLM-5.2
433.8
1,048,576
yes
0.69
13.9
448.3
2.74
14.6
451.2
5.48
15.6
454.9
10.97
17.6
462.4
2026-11-12
glm-5-3
GLM-5.3
Zhipu AI
753B
Q4_K
unsloth/GLM-5.3-GGUF
zai-org/GLM-5.3
435.2
1,048,576
yes
0.69
13.9
449.8
2.74
14.7
452.6
5.48
15.7
456.3
10.97
17.7
463.8
2026-11-12
glm-5-3-flash
GLM-5.3-Flash
Zhipu AI
321B
Q4_K
unsloth/GLM-5.3-Flash-GGUF
zai-org/GLM-5.3-Flash
186
1,048,576
yes
0.09
6.4
192.5
0.34
7.2
193.5
0.69
8.2
194.9
1.38
10.2
197.6
2026-11-12
gpt-oss-120b
gpt-oss 120B
OpenAI
117B MoE
Q4_K_M
unsloth/gpt-oss-120b-GGUF
openai/gpt-oss-120b
58.5
131,072
yes
0.29
2.6
61.4
1.13
3.4
63
2.25
4.4
65.1
4.5
6.4
69.4
2026-11-12
gpt-oss-20b
gpt-oss 20B
OpenAI
20.9B MoE
Q4_K_M
unsloth/gpt-oss-20b-GGUF
openai/gpt-oss-20b
10.8
131,072
yes
0.19
1.2
12.2
0.75
1.9
13.5
1.5
2.9
15.2
3
4.9
18.7
2026-11-12
granite-4-2-30b
Granite 4.2 30B
IBM
29.3B
Q4_K_M
ibm-granite/granite-4.2-30b-GGUF
ibm-granite/granite-4.2-30b
16.5
131,072
yes
2
1.3
19.8
8
2.1
26.6
16
3.1
35.6
32
5.1
53.6
2026-11-12
granite-4-2-3b
Granite 4.2 3B
IBM
3.7B
Q4_K_M
ibm-granite/granite-4.2-3b-GGUF
ibm-granite/granite-4.2-3b
2.1
131,072
yes
0.63
0.9
3.6
2.5
1.7
6.3
5
2.7
9.8
10
4.7
16.8
2026-11-12
granite-4-2-8b
Granite 4.2 8B
IBM
8.8B
Q4_K_M
ibm-granite/granite-4.2-8b-GGUF
ibm-granite/granite-4.2-8b
5
131,072
yes
1.25
1
7.3
5
1.8
11.8
10
2.8
17.8
20
4.8
29.8
2026-11-12
kat-coder-v2-5
KAT-Coder V2.5
Kwaipilot
34.7B MoE
Q4_K_M
bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF
Kwaipilot/KAT-Coder-V2.5-Dev
19.9
262,144
yes
0.16
1.4
21.5
0.63
2.2
22.7
1.25
3.2
24.3
2.5
5.2
27.6
2026-11-12
lfm2-5-1-2b
LFM2.5 1.2B
Liquid AI
1.2B
Q4_K_M
LiquidAI/LFM2.5-1.2B-Instruct-GGUF
LiquidAI/LFM2.5-1.2B-Instruct
0.7
128,000
yes
0.09
0.9
1.7
0.38
1.6
2.7
0.75
2.6
4.1
1.46
4.6
6.8
2026-11-12
lfm2-5-2-6b
LFM2.5 2.6B
Liquid AI
2.7B
Q4_K_M
LiquidAI/LFM2.5-2.6B-GGUF
LiquidAI/LFM2.5-2.6B
1.6
131,072
yes
0.13
0.9
2.6
0.5
1.6
3.7
1
2.6
5.2
2
4.6
8.2
2026-11-12
lfm2-5-24b-a2b
LFM2.5 24B-A2B
Liquid AI
24B MoE
Q4_K_M
LiquidAI/LFM2.5-24B-A2B-GGUF
LiquidAI/LFM2.5-24B-A2B
13.5
131,072
yes
0.16
1.3
14.9
0.63
2
16.1
1.25
3
17.8
2.5
5
21
2026-11-12
lfm2-5-8b-a1b
LFM2.5 8B-A1B
Liquid AI
8.5B MoE
Q4_K_M
LiquidAI/LFM2.5-8B-A1B-GGUF
LiquidAI/LFM2.5-8B-A1B
4.8
128,000
yes
0.09
1
5.9
0.38
1.7
6.9
0.75
2.7
8.3
1.46
4.7
11
2026-11-12
lfm3-1-7b
LFM3 1.7B
Liquid AI
1.7B
Q4_K_M
LiquidAI/LFM3-1.7B-GGUF
LiquidAI/LFM3-1.7B
1
262,144
yes
0.1
0.9
2
0.38
1.6
3
0.75
2.6
4.4
1.5
4.6
7.1
2026-11-12
ling-3-0-flash
Ling 3.0 Flash
InclusionAI
127B MoE
Q4_K_M
AtomicChat/Ling-3.0-flash-GGUF
inclusionAI/Ling-3.0-flash
70.1
262,144
yes
0.37
3
73.4
1.48
3.7
75.3
2.95
4.7
77.8
5.91
6.7
82.7
2026-11-12
ling-3-0-tiny
Ling 3.0 Tiny
InclusionAI
7.9B MoE
Q4_K_M
bloomer010/Ling-3.0-tiny-GGUF
inclusionAI/Ling-3.0-tiny
4.5
131,072
yes
0.21
1
5.7
0.84
1.7
7.1
1.69
2.7
8.9
3.38
4.7
12.6
2026-11-12
llama-3-1-8b
Llama 3.1 8B
Meta
8B
Q4_K_M
unsloth/Llama-3.1-8B-Instruct-GGUF
meta-llama/Llama-3.1-8B-Instruct
4.6
null
no
null
null
null
null
null
null
null
null
null
null
null
null
2026-11-12
llama-3-2-3b
Llama 3.2 3B
Meta
3B
Q4_K_M
unsloth/Llama-3.2-3B-Instruct-GGUF
meta-llama/Llama-3.2-3B-Instruct
1.9
null
no
null
null
null
null
null
null
null
null
null
null
null
null
2026-11-12
llama-3-3-70b
Llama 3.3 70B
Meta
70.6B
Q4_K_M
unsloth/Llama-3.3-70B-Instruct-GGUF
meta-llama/Llama-3.3-70B-Instruct
39.6
131,072
yes
2.5
2
44.1
10
2.8
52.4
20
3.8
63.4
40
5.8
85.4
2026-11-12
mistral-small-3-2
Mistral Small 3.2
Mistral AI
24B
Q4_K_M
unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF
mistralai/Mistral-Small-3.2-24B-Instruct-2506
13.3
131,072
yes
1.25
1.2
15.8
5
2
20.3
10
3
26.3
20
5
38.3
2026-11-12
nemotron-nano-9b
Nemotron Nano 9B v2
NVIDIA
8.9B
Q4_K_M
bartowski/nvidia_NVIDIA-Nemotron-Nano-9B-v2-GGUF
nvidia/NVIDIA-Nemotron-Nano-9B-v2
6.1
131,072
yes
1.75
1
8.9
7
1.8
14.9
14
2.8
22.9
28
4.8
38.9
2026-11-12
ornith-1-5-122b
Ornith 1.5 122B
Ornith AI
122B
Q4_K_M
ornith-ai/Ornith-1.5-122B-GGUF
ornith-ai/Ornith-1.5-122B
70.6
262,144
yes
0.19
3
73.8
0.75
3.7
75.1
1.5
4.7
76.8
3
6.7
80.3
2026-11-12
ornith-1-5-35b
Ornith 1.5 35B-A3B
Ornith AI
36B MoE
Q4_K_M
ornith-ai/Ornith-1.5-35B-A3B-GGUF
ornith-ai/Ornith-1.5-35B-A3B
20.2
262,144
yes
0.16
1.5
21.8
0.63
2.2
23
1.25
3.2
24.7
2.5
5.2
27.9
2026-11-12
ornith-1-5-397b
Ornith 1.5 397B
Ornith AI
403B MoE
Q4_K_M
ornith-ai/Ornith-1.5-397B-GGUF
ornith-ai/Ornith-1.5-397B
227.5
262,144
yes
0.23
7.7
235.4
0.94
8.4
236.9
1.88
9.4
238.8
3.75
11.4
242.7
2026-11-12
ornith-1-5-9b
Ornith 1.5 9B
Ornith AI
9.7B
Q4_K_M
ornith-ai/Ornith-1.5-9B-GGUF
ornith-ai/Ornith-1.5-9B
5.4
262,144
yes
0.25
1
6.7
1
1.8
8.2
2
2.8
10.2
4
4.8
14.2
2026-11-12
phi-4-mini
Phi-4 Mini
Microsoft
3.8B
Q4_K_M
unsloth/Phi-4-mini-instruct-GGUF
microsoft/Phi-4-mini-instruct
2.3
131,072
yes
1
0.9
4.2
4
1.7
8
8
2.7
13
16
4.7
23
2026-11-12
qwen-agentworld-35b
Qwen AgentWorld 35B-A3B
Alibaba Qwen
34.7B MoE
Q4_K_M
unsloth/Qwen-AgentWorld-35B-A3B-GGUF
Qwen/Qwen-AgentWorld-35B-A3B
20.6
262,144
yes
0.16
1.5
22.2
0.63
2.2
23.4
1.25
3.2
25.1
2.5
5.2
28.3
2026-11-12
qwen3-0-6b
Qwen3 0.6B
Alibaba Qwen
0.6B
Q8_0
Qwen/Qwen3-0.6B-GGUF
Qwen/Qwen3-0.6B
0.6
40,960
yes
0.88
0.9
2.3
3.5
1.6
5.7
4.38
2.6
7.6
4.38
4.6
9.6
2026-11-12
qwen3-14b
Qwen3 14B
Alibaba Qwen
14.8B
Q4_K_M
Qwen/Qwen3-14B-GGUF
Qwen/Qwen3-14B
8.4
40,960
yes
1.25
1.1
10.8
5
1.9
15.3
6.25
2.9
17.5
6.25
4.9
19.5
2026-11-12
qwen3-32b
Qwen3 32B
Alibaba Qwen
32.8B
Q4_K_M
Qwen/Qwen3-32B-GGUF
Qwen/Qwen3-32B
18.4
40,960
yes
2
1.4
21.8
8
2.2
28.6
10
3.2
31.6
10
5.2
33.6
2026-11-12
qwen3-4b
Qwen3 4B
Alibaba Qwen
4B
Q4_K_M
unsloth/Qwen3-4B-Instruct-2507-GGUF
Qwen/Qwen3-4B-Instruct-2507
2.3
262,144
yes
1.13
0.9
4.3
4.5
1.7
8.5
9
2.7
14
18
4.7
25
2026-11-12
qwen3-8-2-4t-a95b
Qwen3.8 2.4T-A95B
Alibaba Qwen
2.4T-A95B
IQ4_XS
unsloth/Qwen3.8-2.4T-A95B-GGUF
Qwen/Qwen3.8-2.4T-A95B
1,220.8
262,144
yes
0.72
37.5
1,259
2.88
38.2
1,261.9
5.75
39.2
1,265.8
11.5
41.2
1,273.5
2026-11-12
qwen3-8-27b
Qwen3.8 27B
Alibaba Qwen
27.8B
Q4_K_M
unsloth/Qwen3.8-27B-GGUF
Qwen/Qwen3.8-27B
15.3
262,144
yes
0.5
1.3
17.1
2
2.1
19.4
4
3.1
22.4
8
5.1
28.4
2026-11-12
qwen3-8-flash-next
Qwen3.8 Flash-Next
Alibaba Qwen
180B MoE
Q4_K
unsloth/Qwen3.8-Flash-Next-GGUF
Qwen/Qwen3.8-Flash-Next
103.7
262,144
yes
0.19
4
107.8
0.75
4.7
109.2
1.5
5.7
110.9
3
7.7
114.4
2026-11-12
qwen3-8b
Qwen3 8B
Alibaba Qwen
8.2B
Q4_K_M
Qwen/Qwen3-8B-GGUF
Qwen/Qwen3-8B
4.7
40,960
yes
1.13
1
6.8
4.5
1.7
10.9
5.63
2.7
13.1
5.63
4.7
15.1
2026-11-12
qwen3-coder-30b
Qwen3-Coder 30B-A3B
Alibaba Qwen
30.5B MoE
Q4_K_M
unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF
Qwen/Qwen3-Coder-30B-A3B-Instruct
17.3
262,144
yes
0.75
1.4
19.4
3
2.1
22.4
6
3.1
26.4
12
5.1
34.4
2026-11-12
qwen3-coder-next
Qwen3-Coder Next
Alibaba Qwen
79.7B MoE
Q4_K_M
Qwen/Qwen3-Coder-Next-GGUF
Qwen/Qwen3-Coder-Next
45.1
262,144
yes
0.75
2.2
48.1
3
3
51.1
6
4
55.1
12
6
63.1
2026-11-12
tencent-hy3
Hy3
Tencent
299B MoE
Q4_K_M
DevQuasar/tencent.Hy3-GGUF
tencent/Hy3
166.3
262,144
yes
2.5
5.8
174.6
10
6.6
182.9
20
7.6
193.9
40
9.6
215.9
2026-11-12

Open model VRAM requirements, by context length

How much memory an open model needs to run locally, split into the three things that take it: the file, the context, and the runtime. One row per model, with the memory at four context lengths, for the models tracked on AI Compute Radar.

A second table lists the machines the site tracks, with the memory each can actually spare. Together they answer "does it fit": a model fits a machine at a context length when total_gb_32k <= usable_memory_gb (or the column of another context length).

Files

model_memory.csv

One row per model. The memory columns come four times, once per context length: _8k (8,192 tokens), _32k (32,768), _64k (65,536) and _128k (131,072).

column meaning
model_slug, model_name, creator, parameters the model
quantization the quantization of the measured file, usually Q4_K_M
gguf_repo the Hub repo of the exact GGUF file the size was read from
source_repo the maker's own repo
file_gb measured: size of the published GGUF file
max_context_tokens the model's published window
attention_shape_known no means there is no context figure for this model, on purpose: the site does not guess one
context_cache_gb_* computed: KV cache at that length with an f16 cache, from the model's published attention shape (layers, KV heads, head dimension, sliding windows). Capped at the model's own maximum context. Empty when the shape is not published.
runtime_gb_* assumed: driver context, compute buffer and scratch, 0.6 + 3 % of the file + 0.25 GB per 8K of context
total_gb_* file + context cache + runtime; empty when the context cannot be computed
read_on the day the row was read

hardware.csv

column meaning
hardware_id, name the machine
memory_gb the maker's figure
usable_memory_gb what a model can be given: 88 % of the memory on a plug-in card (a 12 % safety margin for the display, the compositor and fragmentation), and the share macOS wires to the GPU on a Mac
memory_architecture discrete (a card with its own memory) or unified
category, page the class, and the machine's page with every model checked against it

What these numbers are, and are not

  • File sizes are measured. Context memory is computed. The runtime share and the safety margin are assumptions, stated above.
  • They are estimates of memory, built for llama.cpp-style runtimes (llama.cpp, Ollama, LM Studio) loading a GGUF file. Real use moves with batch size, flash attention and runtime version. A quantized KV cache (q8_0, q4_0) needs less than the f16 figure here.
  • There is no speed in this dataset: no tokens per second, no benchmark. The site does not measure them.
  • Figures are rounded as the API serves them, so the three parts can differ from the total by up to 0.1 GB.
  • A snapshot, dated in read_on. The live numbers are on the site and in its API.

Use

import pandas as pd

memory = pd.read_csv("hf://datasets/CyferUnknown/open-model-vram-requirements/model_memory.csv")
hardware = pd.read_csv("hf://datasets/CyferUnknown/open-model-vram-requirements/hardware.csv")

card = hardware.set_index("hardware_id").loc["rtx-4090"]
fits_at_32k = memory[memory.total_gb_32k <= card.usable_memory_gb]
print(fits_at_32k[["model_name", "file_gb", "total_gb_32k"]])

Source, licence, citation

Built from the public API of AI Compute Radar (https://aicomputeradar.dev/developers), which reads model metadata and file sizes from the Hugging Face Hub. Method: https://aicomputeradar.dev/methodology.

Licence: CC BY 4.0. Please cite as:

AI Compute Radar (2026). Open model VRAM requirements by context length. https://aicomputeradar.dev

See it as a picture: the card, seen through.

Downloads last month
55

Space using CyferUnknown/open-model-vram-requirements 1