DavidAU · text

ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking

DavidAU/ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking

ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking at Q4_K_M is exactly 13,245,832,896 bytes (12.34 GiB / 13.25 GB) — an effective 4.855 bits per weight, not the nominal 4. Its KV cache at 32K is 1.75 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
21.8B
Architecture
ernie4_5-moe
28 layers
Context
131,072
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
I1-IQ1_S4.17 GiB4,476,481,5361.641—mradermacher
I1-IQ1_M4.63 GiB4,969,721,8561.822—mradermacher
I1-IQ2_XXS5.39 GiB5,791,789,0562.123—mradermacher
I1-IQ2_XS6.01 GiB6,455,177,2162.366—mradermacher
I1-IQ2_S6.06 GiB6,510,534,6562.386—mradermacher
I1-IQ2_M6.68 GiB7,168,188,4162.627—mradermacher
I1-Q2_K_S6.94 GiB7,448,600,5762.730—mradermacher
Q2_K7.50 GiB8,053,067,4562.952—mradermacher
I1-Q2_K7.50 GiB8,053,067,7762.952—mradermacher
I1-IQ3_XXS7.88 GiB8,456,093,6963.099—mradermacher
I1-IQ3_XS8.37 GiB8,983,883,7763.293—mradermacher
Q3_K_S8.85 GiB9,500,266,1763.482—mradermacher
I1-Q3_K_S8.85 GiB9,500,266,4963.482—mradermacher
I1-IQ3_S8.85 GiB9,505,140,7363.484—mradermacher
I1-IQ3_M8.94 GiB9,602,625,5363.520—mradermacher
Q3_K_M9.75 GiB10,468,581,0563.837—mradermacher
I1-Q3_K_M9.75 GiB10,468,581,3763.837—mradermacher
Q3_K_L10.59 GiB11,371,667,1364.168—mradermacher
I1-Q3_K_L10.59 GiB11,371,667,4564.168—mradermacher
I1-IQ4_XS10.89 GiB11,695,292,4164.287—mradermacher
IQ4_XS10.97 GiB11,781,308,0964.318—mradermacher
I1-Q4_011.55 GiB12,397,592,5764.544—mradermacher
Q4_K_S11.58 GiB12,432,654,0164.557—mradermacher
I1-Q4_K_S11.58 GiB12,432,654,3364.557—mradermacher
Q4_K_M12.34 GiB13,245,832,8964.855—mradermacher
I1-Q4_K_M12.34 GiB13,245,833,2164.855—mradermacher
I1-Q4_112.77 GiB13,710,442,4965.026—mradermacher
Q5_K_S14.02 GiB15,057,698,4965.519—mradermacher
I1-Q5_K_S14.02 GiB15,057,698,8165.519—mradermacher
Q5_K_M14.45 GiB15,512,395,4565.686—mradermacher
I1-Q5_K_M14.45 GiB15,512,395,7765.686—mradermacher
Q6_K16.69 GiB17,920,618,1766.569—mradermacher
I1-Q6_K16.69 GiB17,920,618,4966.569—mradermacher
Q8_021.61 GiB23,205,359,2968.506—mradermacher

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.22 GiB0.22 GiB—28 / 0 / 0
8,1920.44 GiB0.44 GiB—28 / 0 / 0
16,3840.88 GiB0.88 GiB—28 / 0 / 0
32,7681.75 GiB1.75 GiB—28 / 0 / 0
65,5363.50 GiB3.50 GiB—28 / 0 / 0
131,0727.00 GiB7.00 GiB—28 / 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 11.43 GiB. The real file is 12.34 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
28
Attention heads
20
KV heads
4
Head dim
128
Hidden size
2560
Vocab
103,424
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking need?
Q4_K_M is exactly 13,245,832,896 bytes (12.34 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking's KV cache?
1.75 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 ERNIE-21B-A3B-Claude-4.5-High-OPUS-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.