Model

OpenHermes 2.5 Mistral 7B

OPEN

Nous Research · OpenHermes family · released Nov 3, 2023

Hugely influential Mistral-7B fine-tune on ~1M mostly GPT-4-generated samples; adding code data lifted HumanEval from 43% to 50.7% pass@1.

ReasoningCodingVisionFunction callingTool useAgentic

186.9

Elo · rank #373

Parameters

7.24B

Active params

7.24B (dense)

Context

8K tokens

Architecture

Dense transformer decoder (Mistral architecture: GQA + sliding-window attention)

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

AGIEvalReasoning43.1%#27

best: OLMo 3-Think 32B · 88.2%

ARC-ChallengeReasoning60.1%#89

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning81.7%#21

best: Phi-3-medium (14B) · 97.7%

BIG-benchReasoning41.0%#5

best: Chinchilla · 65.1%

DROPReasoning35.79#59

best: Hunyuan-T1 · 93.1

GSM8KMath26.1%#159

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning81.7%#69

best: Claude 3 Opus · 95.4%

HumanEvalCoding50.7%#121

best: Claude Opus 4.5 · 99.4%

MMLUKnowledge63.6%#186

best: OpenAI o3 · 92.9%

OpenBookQAReasoning44.8%#37

best: Claude 1 · 90.8%

PIQAReasoning82.7%#22

best: GPT-4o mini · 93.1%

TruthfulQAKnowledge53.0%#45

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

WinoGrandeReasoning74.3%#72

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

6 GB

VRAM @ FP16

15 GB

Fits on (Q4)

RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

Throughput data unavailable.

Quantizations

GGUF Q4_K_M · GPTQ · AWQ · MLX

Fine-tune it

Permissive

QLoRA6.6 GB1× RTX 3060 12GB

LoRA17.1 GB1× RTX 3090 24GB

Full fine-tune117.8 GB1× H200 141GB

QLoRA SFT on ~10k samples ≈ $4.24 (1× RTX 3060 12GB)

Plan a fine-tune →

OpenHermes family

Elo progression across releases

OpenHermes 2.5 Mistral 7BNov 2023186.9

Resources

HFteknium/OpenHermes-2.5-Mistral-7B APIAvailable via Nous Research API

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