Model

Hermes 3 405B

OPEN

Nous Research · Hermes family · released Aug 15, 2024

First full-parameter fine-tune of Llama 3.1 405B; Nous's neutrally-aligned flagship generalist with strong agentic, function-calling and roleplay ability.

ReasoningCodingVisionFunction callingTool useAgentic

842.1

Elo · rank #247

Parameters

405B

Active params

405B (dense)

Context

128K tokens

Architecture

Dense transformer decoder (Llama 3.1 architecture, grouped-query attention)

License

Llama 3.1 Community License

Languages

API price (in/out)

$1 / $1

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AGIEvalReasoning61.8%#8

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning69.5%#50

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning86.2%#12

best: Phi-3-medium (14B) · 97.7%

BIG-Bench HardReasoning75.4%#40

best: ERNIE 4.5 300B-A47B · 94.3%

GPQA DiamondReasoning44.8%#206

best: GPT-5.6 · 94.6%

HellaSwagReasoning90.2%#8

best: Claude 3 Opus · 95.4%

IFEvalReasoning84.9%#62

best: Gemma 4 26B A4B · 98.5%

MATH-500Math30.9%#163

best: GPT-5 · 99.4%

MMLU-ProKnowledge54.1%#126

best: Claude Fable 5 · 91.5%

MMLUKnowledge85.0%#59

best: OpenAI o3 · 92.9%

MT-BenchHuman preference8.93#14

best: Hunyuan-Large (A52B) · 9.4

OpenBookQAReasoning48.8%#27

best: Claude 1 · 90.8%

PIQAReasoning86.0%#8

best: GPT-4o mini · 93.1%

TruthfulQAKnowledge65.6%#14

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WinoGrandeReasoning86.3%#9

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

250 GB

VRAM @ FP16

810 GB

Fits on (Q4)

M3 Ultra 512GB

Throughput data unavailable.

Quantizations

FP8 · GGUF Q4_K_M

Hermes family

Elo progression across releases

Nous Hermes 2 Mixtral 8x7B DPOJan 2024474.1 Hermes 3 405BAug 2024842.1

Resources

HFNousResearch/Hermes-3-Llama-3.1-405B DOCSnousresearch.com/hermes3 APIAvailable via Nous Research API

API price $1/$1 · each benchmark row carries its own source badge (see methodology)