Model

Nous Hermes 2 Mixtral 8x7B DPO

OPEN

Nous Research · Hermes family · released Jan 15, 2024

Flagship MoE fine-tune of Mixtral 8x7B, SFT+DPO on 1M+ mostly GPT-4-generated samples; first community model to beat Mixtral Instruct broadly.

ReasoningCodingVisionFunction callingTool useAgentic

474.1

Elo · rank #314

Parameters

46.7B

Active params

12.9B (MoE)

Context

32K tokens

Architecture

Sparse Mixture-of-Experts transformer (8 experts, top-2 routing, Mixtral architecture)

License

Apache-2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

AGIEvalReasoning46.0%#24

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning64.3%#68

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning86.4%#11

best: Phi-3-medium (14B) · 97.7%

BIG-benchReasoning49.7%#3

best: Chinchilla · 65.1%

GSM8KMath71.7%#98

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning84.9%#43

best: Claude 3 Opus · 95.4%

MMLUKnowledge72.2%#135

best: OpenAI o3 · 92.9%

OpenBookQAReasoning46.6%#33

best: Claude 1 · 90.8%

PIQAReasoning83.8%#13

best: GPT-4o mini · 93.1%

TruthfulQAKnowledge54.8%#37

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

WinoGrandeReasoning76.2%#56

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

30 GB

VRAM @ FP16

93 GB

Fits on (Q4)

M4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4_K_M · GPTQ · AWQ

Hermes family

Elo progression across releases

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

Resources

HFNousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO APIAvailable via Nous Research API

API price weights · each benchmark row carries its own source badge (see methodology)