NousResearch · text

Hermes-4-405B

NousResearch/Hermes-4-405B
Editorial review at Uncensor Index →

Hermes-4-405B at Q4_K_M is exactly 243,069,612,160 bytes (226.38 GiB / 243.07 GB) — an effective 4.791 bits per weight, not the nominal 4. Its KV cache at 32K is 15.75 GiB.

From the file· summed from 5 file(s)From the file· KV per layer
Parameters
406B
Architecture
llama
126 layers
Context
131,072
native (config.json)
License
llama3

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
UD-IQ1_S2 shards81.23 GiB87,216,287,4561.719—unsloth
IQ1_M7 shards87.08 GiB93,498,011,3601.843—DevQuasar
UD-IQ1_M2 shards88.23 GiB94,734,053,1201.867—unsloth
IQ2_XXS8 shards99.91 GiB107,272,368,0322.115—DevQuasar
IQ2_XS9 shards111.15 GiB119,348,817,9202.353—DevQuasar
IQ2_S9 shards117.02 GiB125,647,646,7202.477—DevQuasar
Q2_K11 shards139.07 GiB149,322,788,8962.943—DevQuasar
Q3_K_M15 shards181.96 GiB195,373,396,4163.851—DevQuasar
Q4_K_M5 shards226.38 GiB243,069,612,1604.791—unsloth
Q4_K_M18 shards226.38 GiB243,069,612,7684.791—DevQuasar
Q5_K_M22 shards266.96 GiB286,646,859,0405.650—DevQuasar
Q6_K25 shards310.08 GiB332,947,682,9126.563—DevQuasar
Q8_09 shards401.62 GiB431,239,295,3288.500—unsloth
Q8_032 shards401.62 GiB431,239,297,4408.500—DevQuasar
BF1617 shards755.98 GiB811,722,962,24016.000—unsloth
F1655 shards755.98 GiB811,722,966,20816.000—DevQuasar

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0961.97 GiB1.97 GiB—126 / 0 / 0
8,1923.94 GiB3.94 GiB—126 / 0 / 0
16,3847.88 GiB7.88 GiB—126 / 0 / 0
32,76815.75 GiB15.75 GiB—126 / 0 / 0
65,53631.50 GiB31.50 GiB—126 / 0 / 0
131,07263.00 GiB63.00 GiB—126 / 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 212.61 GiB. The real file is 226.38 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
126
Attention heads
128
KV heads
8
Head dim
128
Hidden size
16384
Vocab
128,256
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Hermes-4-405B need?
Q4_K_M is exactly 243,069,612,160 bytes (226.38 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Hermes-4-405B's KV cache?
15.75 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 Hermes-4-405B 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.