huihui-ai · text

Huihui-Qwen3-VL-32B-Instruct-abliterated

huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated

Huihui-Qwen3-VL-32B-Instruct-abliterated at Q4_K_M is exactly 19,762,151,264 bytes (18.40 GiB / 19.76 GB) — an effective 4.739 bits per weight, not the nominal 4. Its KV cache at 32K is 8.00 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
33.4B
Architecture
qwen3vl
64 layers
Context
262,144
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
I1-IQ1_S6.82 GiB7,323,843,7121.756—mradermacher
I1-IQ1_M7.41 GiB7,959,870,5921.909—mradermacher
I1-IQ2_XXS8.40 GiB9,019,915,3922.163—mradermacher
I1-IQ2_XS9.27 GiB9,951,837,3122.387—mradermacher
I1-IQ2_S9.79 GiB10,514,827,3922.522—mradermacher
I1-IQ2_M10.58 GiB11,362,863,2322.725—mradermacher
I1-Q2_K_S10.68 GiB11,465,816,1922.750—mradermacher
Q2_K11.50 GiB12,344,653,6642.961—mradermacher
I1-Q2_K11.50 GiB12,344,653,9522.961—mradermacher
I1-IQ3_XXS11.94 GiB12,821,039,2323.075—mradermacher
I1-IQ3_XS12.76 GiB13,702,923,3923.286—mradermacher
Q3_K_S13.40 GiB14,389,740,3843.451—mradermacher
I1-Q3_K_S13.40 GiB14,389,740,6723.451—mradermacher
I1-IQ3_S13.44 GiB14,434,305,1523.462—mradermacher
I1-IQ3_M13.90 GiB14,930,084,9923.581—mradermacher
Q3_K_M14.87 GiB15,971,779,4243.830—mradermacher
I1-Q3_K_M14.87 GiB15,971,779,7123.830—mradermacher
Q3_K_L16.14 GiB17,330,996,0644.156—mradermacher
I1-Q3_K_L16.14 GiB17,330,996,3524.156—mradermacher
I1-IQ4_XS16.48 GiB17,690,497,1524.243—mradermacher
IQ4_XS16.63 GiB17,854,336,8644.282—mradermacher
I1-Q4_017.42 GiB18,703,089,7924.486—mradermacher
Q4_K_S17.48 GiB18,771,246,9444.502—mradermacher
I1-Q4_K_S17.48 GiB18,771,247,2324.502—mradermacher
Q4_K_M18.40 GiB19,762,151,2644.739—mradermacher
I1-Q4_K_M18.40 GiB19,762,151,5524.739—mradermacher
I1-Q4_119.22 GiB20,636,524,6724.949—mradermacher
Q5_K_S21.08 GiB22,635,495,2645.429—mradermacher
I1-Q5_K_S21.08 GiB22,635,495,5525.429—mradermacher
Q5_K_M21.62 GiB23,214,833,5045.567—mradermacher
I1-Q5_K_M21.62 GiB23,214,833,7925.567—mradermacher
Q6_K25.04 GiB26,883,308,3846.447—mradermacher
I1-Q6_K25.04 GiB26,883,308,6726.447—mradermacher
Q8_032.43 GiB34,817,721,1848.350—mradermacher

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0961.00 GiB1.00 GiB—64 / 0 / 0
8,1922.00 GiB2.00 GiB—64 / 0 / 0
16,3844.00 GiB4.00 GiB—64 / 0 / 0
32,7688.00 GiB8.00 GiB—64 / 0 / 0
65,53616.00 GiB16.00 GiB—64 / 0 / 0
131,07232.00 GiB32.00 GiB—64 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-12B-it-qat-q4_0-unquantizedvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507All 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 17.47 GiB. The real file is 18.40 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
64
Attention heads
64
KV heads
8
Head dim
128
Hidden size
5120
Vocab
151,936
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Huihui-Qwen3-VL-32B-Instruct-abliterated need?
Q4_K_M is exactly 19,762,151,264 bytes (18.40 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Huihui-Qwen3-VL-32B-Instruct-abliterated's KV cache?
8.00 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.
Which quantization of Huihui-Qwen3-VL-32B-Instruct-abliterated 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.