Model

H2O-Danube3 4B

OPEN

H2O.ai · H2O-Danube family · released Jul 12, 2024

Apache-2.0 small model trained on 6T tokens, compact enough to run locally on a modern smartphone.

ReasoningCodingVisionFunction callingTool useAgentic

213.1

Elo · rank #360

Parameters

3.96B

Active params

3.96B (dense)

Context

8K tokens

Architecture

Dense Llama-2-style transformer, wide architecture, GQA (24 layers, 3840 hidden)

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

ARC-ChallengeReasoning59.0%#96

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning83.8%#15

best: Phi-3-medium (14B) · 97.7%

BIG-Bench HardReasoning38.9%#115

best: ERNIE 4.5 300B-A47B · 94.3%

GSM8KMath50.2%#134

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning80.4%#78

best: Claude 3 Opus · 95.4%

MMLUKnowledge54.7%#223

best: OpenAI o3 · 92.9%

MT-BenchHuman preference6.49#62

best: Hunyuan-Large (A52B) · 9.4

PIQAReasoning82.6%#23

best: GPT-4o mini · 93.1%

TruthfulQAKnowledge47.8%#60

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

WinoGrandeReasoning76.5%#54

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

2.6 GB

VRAM @ FP16

8.3 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

H2O-Danube family

Elo progression across releases

H2O-Danube 1.8BJan 2024-285.2 H2O-Danube3 4BJul 2024213.1

Resources

HFh2oai/h2o-danube3-4b-chat DOCSh2o.ai/platform/danube

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