Qwen · text

Qwen2.5-VL-3B-Instruct

Qwen/Qwen2.5-VL-3B-Instruct

Qwen2.5-VL-3B-Instruct at Q4_K_M is exactly 1,929,901,056 bytes (1.80 GiB / 1.93 GB) — an effective 4.112 bits per weight, not the nominal 4. Its KV cache at 32K is 1.13 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
3.8B
Architecture
qwen2vl
36 layers
Context
128,000
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
UD-IQ1_S0.77 GiB826,334,5601.761—unsloth
UD-IQ1_M0.82 GiB881,638,7521.879—unsloth
UD-IQ2_XXS0.90 GiB970,685,7922.068—unsloth
UD-IQ2_M1.09 GiB1,165,401,4402.483—unsloth
Q2_K_L1.19 GiB1,274,754,4002.716—unsloth
Q2_K1.19 GiB1,274,754,4002.716—unsloth
UD-IQ3_XXS1.21 GiB1,302,654,3042.776—unsloth
Q3_K_S1.35 GiB1,454,355,8083.099—unsloth
Q3_K_M1.48 GiB1,590,474,0803.389434unsloth
Q3_K_L1.59 GiB1,707,389,9523.638—lmstudio-community
IQ4_XS1.62 GiB1,739,093,3443.705434unsloth
IQ4_NL1.70 GiB1,825,207,6483.889—unsloth
Q4_01.70 GiB1,828,484,4483.896434unsloth
Q4_K_S1.71 GiB1,834,382,6883.909—unsloth
Q4_K_M1.80 GiB1,929,901,0564.112—ggml-org
Q4_K_M1.80 GiB1,929,901,0564.112—lmstudio-community
Q4_K_M1.80 GiB1,929,901,4084.112434unsloth
Q4_11.86 GiB1,996,256,6084.253—unsloth
Q5_K_S2.02 GiB2,169,664,8644.623—unsloth
Q5_K_M2.07 GiB2,224,813,4084.740434unsloth
Q6_K2.36 GiB2,538,157,0565.408434lmstudio-community
Q6_K2.36 GiB2,538,157,4085.408—unsloth
Q8_03.06 GiB3,285,474,3047.000—lmstudio-community
Q8_03.06 GiB3,285,474,3047.000434ggml-org
Q8_03.06 GiB3,285,474,6567.000—unsloth
F165.75 GiB6,178,315,26413.164434ggml-org
BF165.75 GiB6,178,315,32813.164—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.14 GiB0.14 GiB—36 / 0 / 0
8,1920.28 GiB0.28 GiB—36 / 0 / 0
16,3840.56 GiB0.56 GiB—36 / 0 / 0
32,7681.13 GiB1.13 GiB—36 / 0 / 0
65,5362.25 GiB2.25 GiB—36 / 0 / 0
131,0724.50 GiB4.50 GiB—36 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-4Bvs HyperCLOVAX-SEED-Text-Instruct-1.5Bvs Llama-3.2-1B-Instructvs ced-basevs gemma-3-1b-itvs Qwen3-1.7BAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 1.97 GiB. The real file is 1.80 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
36
Attention heads
16
KV heads
2
Head dim
128
Hidden size
2048
Vocab
151,936
Sliding window
32768
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
false

This model declares a sliding window but sets use_sliding_window: false, so the window is not applied. Honouring the field without the flag understates KV for the whole family.

Questions people ask

How much VRAM does Qwen2.5-VL-3B-Instruct need?
Q4_K_M is exactly 1,929,901,056 bytes (1.80 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen2.5-VL-3B-Instruct's KV cache?
1.13 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.
Which quantization of Qwen2.5-VL-3B-Instruct 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.