Model

H2O-Danube 1.8B

OPEN

H2O.ai · H2O-Danube family · released Jan 30, 2024

H2O.ai's debut small LLM, Llama-style architecture, tuned for efficient on-device/mobile inference.

ReasoningCodingVisionFunction callingTool useAgentic

-339.0

Elo · rank #450

Parameters

1.8B

Active params

1.8B (dense)

Context

16K tokens

Architecture

Dense Llama-2/Mistral-style transformer with sliding-window attention + GQA

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-ChallengeReasoning39.3%#150

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning67.5%#54

best: Phi-3-medium (14B) · 97.7%

GSM8KMath15.5%#169

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning67.6%#133

best: Claude 3 Opus · 95.4%

MMLUKnowledge33.5%#264

best: OpenAI o3 · 92.9%

MT-BenchHuman preference5.52#66

best: Hunyuan-Large (A52B) · 9.4

OpenBookQAReasoning39.2%#45

best: Claude 1 · 90.8%

PIQAReasoning76.7%#64

best: GPT-4o mini · 93.1%

TriviaQAKnowledge36.3%#47

best: Sarvam-1 (2B) · 90.6%

TruthfulQAKnowledge40.8%#78

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

WinoGrandeReasoning65.3%#121

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

1.4 GB

VRAM @ FP16

3.8 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

Fine-tune it

Permissive

QLoRA3.1 GB1× RTX 3060 12GB

LoRA5.7 GB1× RTX 3060 12GB

Full fine-tune30.8 GB1× RTX 5090 32GB

QLoRA SFT on ~10k samples ≈ $1.05 (1× RTX 3060 12GB)

Plan a fine-tune →

H2O-Danube family

Elo progression across releases

H2O-Danube 1.8BJan 2024-339.0 H2O-Danube3 4BJul 2024124.0

Resources

HFh2oai/h2o-danube-1.8b-chat DOCSarxiv.org/abs/2401.16818

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