NousResearch · text

Hermes-4-70B

NousResearch/Hermes-4-70B

Hermes-4-70B at Q4_K_M is exactly 42,520,397,472 bytes (39.60 GiB / 42.52 GB) — an effective 4.821 bits per weight, not the nominal 4. Its KV cache at 32K is 10.00 GiB.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
70.6B
Architecture
llama
80 layers
Context
131,072
native (config.json)
License
—

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
UD-IQ1_S14.77 GiB15,854,029,9841.798—unsloth
IQ1_M15.60 GiB16,751,200,8001.899—bartowski
UD-IQ1_M15.97 GiB17,145,875,6161.944—unsloth
IQ2_XXS17.79 GiB19,097,389,6002.165—bartowski
UD-IQ2_XXS18.10 GiB19,431,115,9362.203—unsloth
IQ2_XS19.69 GiB21,142,112,8002.397—bartowski
IQ2_S20.71 GiB22,242,347,5522.522—bartowski
IQ2_M22.46 GiB24,119,298,5922.735—bartowski
UD-IQ2_M22.63 GiB24,295,197,8562.755—unsloth
Q2_K24.56 GiB26,375,113,2482.991—bartowski
Q2_K24.56 GiB26,375,113,8882.991—unsloth
Q2_K_L24.79 GiB26,621,365,4083.019—unsloth
Q2_K_L25.52 GiB27,401,161,2483.107—bartowski
IQ3_XXS25.58 GiB27,469,498,9123.115—bartowski
UD-IQ3_XXS25.76 GiB27,655,621,7923.136—unsloth
IQ3_XS27.29 GiB29,307,734,5603.323—bartowski
Q3_K_S28.79 GiB30,912,055,8403.505—bartowski
Q3_K_S28.79 GiB30,912,056,4803.505—unsloth
IQ3_M29.74 GiB31,937,038,8803.621—bartowski
Q3_K_M31.91 GiB34,267,499,0403.886—bartowski
Q3_K_M31.91 GiB34,267,499,6803.886—unsloth
Q3_K_L34.59 GiB37,140,596,3844.211—lmstudio-community
Q3_K_L34.59 GiB37,140,597,2804.211—bartowski
IQ4_XS35.30 GiB37,902,666,2724.298—bartowski
IQ4_XS35.33 GiB37,935,500,4484.301—unsloth
IQ4_NL37.30 GiB40,053,623,3284.542—bartowski
IQ4_NL37.30 GiB40,053,623,9684.542—unsloth
Q4_037.36 GiB40,116,537,8884.549—bartowski
Q4_037.36 GiB40,116,538,5284.549—unsloth
Q4_K_S37.58 GiB40,347,224,6084.575—bartowski
Q4_K_S37.58 GiB40,347,225,2484.575—unsloth
Q4_K_M39.60 GiB42,520,397,4724.821—lmstudio-community
Q4_K_M39.60 GiB42,520,398,3684.821—bartowski
Q4_K_M39.60 GiB42,520,399,0084.821—unsloth
Q4_141.27 GiB44,313,594,4005.025—bartowski
Q4_141.27 GiB44,313,595,0405.025—unsloth
Q5_K_S45.32 GiB48,657,451,5525.517—bartowski
Q5_K_S45.32 GiB48,657,452,1925.517—unsloth
Q5_K_M2 shards46.52 GiB49,949,821,6965.664—bartowski
Q5_K_M46.52 GiB49,949,822,1125.664—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0961.25 GiB1.25 GiB—80 / 0 / 0
8,1922.50 GiB2.50 GiB—80 / 0 / 0
16,3845.00 GiB5.00 GiB—80 / 0 / 0
32,76810.00 GiB10.00 GiB—80 / 0 / 0
65,53620.00 GiB20.00 GiB—80 / 0 / 0
131,07240.00 GiB40.00 GiB—80 / 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 36.96 GiB. The real file is 39.60 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
80
Attention heads
64
KV heads
8
Head dim
128
Hidden size
8192
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-4-70B need?
Q4_K_M is exactly 42,520,397,472 bytes (39.60 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-4-70B's KV cache?
10.00 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 Hermes-4-70B 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.