empero-ai · vision language

Qwythos-9B-Claude-Mythos-5-1M

empero-ai/Qwythos-9B-Claude-Mythos-5-1M

Qwythos-9B-Claude-Mythos-5-1M at Q4_K_M is exactly 11,516,776,704 bytes (10.73 GiB / 11.52 GB) — an effective 9.791 bits per weight, not the nominal 4. Its KV cache at 32K is 1.00 GiB.

From the file· summed from 2 file(s)From the file· KV per layer
Parameters
9.4B
Architecture
qwen35
32 layers
Context
1,048,576
native (config.json)
License
apache-2.0

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
Q4_K5.38 GiB5,780,091,0404.914—huihui-ai
Q5_K6.19 GiB6,642,544,8005.647—huihui-ai
Q6_K7.04 GiB7,558,901,9206.426442huihui-ai
Q8_09.11 GiB9,786,060,9608.320442huihui-ai
Q4_K_M2 shards10.73 GiB11,516,776,7049.791—empero-ai
Q4_K_M2 shards10.73 GiB11,516,777,4089.791—deewu0809
Q5_K_M2 shards12.25 GiB13,152,614,78411.182—wepiqx
Q5_K_M2 shards12.29 GiB13,194,498,30411.218—empero-ai
Q5_K_M2 shards12.29 GiB13,194,499,00811.218—deewu0809
Q6_K2 shards13.95 GiB14,977,077,50412.733—empero-ai
Q6_K2 shards13.95 GiB14,977,078,20812.733—deewu0809
BF1617.14 GiB18,407,321,76015.649—huihui-ai
Q8_02 shards17.99 GiB19,313,561,85616.420—empero-ai
Q8_02 shards17.99 GiB19,313,562,56016.420—deewu0809
BF162 shards33.83 GiB36,328,018,17630.885—empero-ai
BF162 shards33.83 GiB36,328,018,88030.885—deewu0809

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0960.13 GiB0.50 GiB4.00×8 / 0 / 24
8,1920.25 GiB1.00 GiB4.00×8 / 0 / 24
16,3840.50 GiB2.00 GiB4.00×8 / 0 / 24
32,7681.00 GiB4.00 GiB4.00×8 / 0 / 24
65,5362.00 GiB8.00 GiB4.00×8 / 0 / 24
131,0724.00 GiB16.00 GiB4.00×8 / 0 / 24

24 of 32 layers use linear attention, which keeps a fixed-size recurrent state instead of a per-token cache. Those layers do not grow with context at all — treating them as ordinary attention, as a flat formula does, overstates this model's cache by roughly 4.0× at long context.

Compare with

same modality, comparable size
vs Qwen3.6-35B-A3Bvs Qwen3.5-9Bvs gemma-4-26B-A4B-itvs Qwen3.5-4Bvs Qwen3-VL-30B-A3B-Instructvs gemma-4-26B-A4B-it-qat-q4_0-unquantizedAll 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 32GBApple M5 Max 36GBApple M5 Max 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 4.93 GiB. The real file is 10.73 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
32
Attention heads
16
KV heads
4
Head dim
256
Hidden size
4096
Vocab
248,320
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Qwythos-9B-Claude-Mythos-5-1M need?
Q4_K_M is exactly 11,516,776,704 bytes (10.73 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Qwythos-9B-Claude-Mythos-5-1M's KV cache?
1.00 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 Qwythos-9B-Claude-Mythos-5-1M 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.