Qwen · text · mixture of experts

Qwen3-VL-235B-A22B-Thinking

Qwen/Qwen3-VL-235B-A22B-Thinking

Qwen3-VL-235B-A22B-Thinking at Q4_K_M is exactly 142,154,075,360 bytes (132.39 GiB / 142.15 GB) — an effective 4.825 bits per weight, not the nominal 4. Its KV cache at 32K is 5.88 GiB.

From the file· summed from 4 file(s)From the file· KV per layer
Parameters
236B
total, not active
Architecture
qwen3vlmoe
94 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
UD-TQ1_050.83 GiB54,574,584,0321.853—unsloth
UD-TQ1_050.92 GiB54,675,246,7201.856—unsloth
UD-IQ1_S2 shards58.57 GiB62,891,889,0562.135—unsloth
UD-IQ1_S2 shards58.65 GiB62,973,677,3762.138—unsloth
UD-IQ1_M2 shards64.83 GiB69,608,509,8562.363—unsloth
UD-IQ1_M2 shards64.90 GiB69,690,298,1762.366—unsloth
Q2_K2 shards79.81 GiB85,691,002,6882.909—unsloth
Q2_K2 shards79.81 GiB85,691,003,3282.909—unsloth
Q2_K_L2 shards79.94 GiB85,836,861,2482.914—unsloth
Q2_K_L2 shards79.94 GiB85,836,861,8562.914—unsloth
UD-IQ3_XXS2 shards88.15 GiB94,648,504,7683.213—unsloth
UD-IQ3_XXS2 shards88.22 GiB94,730,293,0563.216—unsloth
Q3_K_S3 shards94.48 GiB101,444,706,2403.444—unsloth
Q3_K_S3 shards94.48 GiB101,444,706,8803.444—unsloth
Q3_K_M3 shards104.72 GiB112,447,381,4083.817—unsloth
Q3_K_M3 shards104.72 GiB112,447,382,0483.817—unsloth
Q8_03 shards116.38 GiB124,966,771,3924.242—unsloth
IQ4_XS3 shards116.70 GiB125,303,361,5044.253—unsloth
IQ4_XS3 shards116.70 GiB125,303,362,1124.253—unsloth
IQ4_NL3 shards123.50 GiB132,603,416,4804.501—unsloth
IQ4_NL3 shards123.50 GiB132,603,417,1204.501—unsloth
Q4_03 shards123.99 GiB133,132,423,1044.519—unsloth
Q4_03 shards123.99 GiB133,132,423,7124.519—unsloth
Q4_K_S3 shards124.51 GiB133,687,119,8084.538—unsloth
Q4_K_S3 shards124.51 GiB133,687,120,4164.538—unsloth
Q5_K_S3 shards125.69 GiB134,959,552,0644.581—unsloth
Q4_K_M4 shards132.39 GiB142,154,075,3604.825—lmstudio-community
Q4_K_M3 shards132.39 GiB142,154,076,0964.825—unsloth
Q4_K_M3 shards132.39 GiB142,154,076,7044.825—unsloth
Q4_13 shards137.12 GiB147,230,101,4404.998—unsloth
Q4_13 shards137.12 GiB147,230,102,0804.998—unsloth
Q5_K_S4 shards150.76 GiB161,881,428,0325.495—unsloth
Q5_K_M4 shards155.36 GiB166,814,158,8485.663—unsloth
Q5_K_M4 shards155.36 GiB166,814,159,4885.663—unsloth
Q6_K5 shards179.76 GiB193,015,495,9686.552—lmstudio-community
Q6_K4 shards179.76 GiB193,015,496,7686.552—unsloth
Q6_K4 shards179.76 GiB193,015,497,3766.552—unsloth
Q8_06 shards232.77 GiB249,940,106,6568.484—Qwen
Q8_07 shards232.77 GiB249,940,106,8488.484—lmstudio-community
Q8_06 shards232.77 GiB249,940,107,5848.484—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.73 GiB0.73 GiB—94 / 0 / 0
8,1921.47 GiB1.47 GiB—94 / 0 / 0
16,3842.94 GiB2.94 GiB—94 / 0 / 0
32,7685.88 GiB5.88 GiB—94 / 0 / 0
65,53611.75 GiB11.75 GiB—94 / 0 / 0
131,07223.50 GiB23.50 GiB—94 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs DeepSeek-V4-Flashvs Qwen3-30B-A3B-Thinking-2507vs Laguna-XS-2.1All 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 123.46 GiB. The real file is 132.39 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
94
Attention heads
64
KV heads
4
Head dim
128
Hidden size
4096
Vocab
151,936
Sliding window
none
SWA period
—
MLA
no
Experts
128
Experts per token
8
use_sliding_window
—

Questions people ask

How much VRAM does Qwen3-VL-235B-A22B-Thinking need?
Q4_K_M is exactly 142,154,075,360 bytes (132.39 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwen3-VL-235B-A22B-Thinking's KV cache?
5.88 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.
Is Qwen3-VL-235B-A22B-Thinking a mixture-of-experts model?
Yes — 128 experts, 8 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of Qwen3-VL-235B-A22B-Thinking 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.