Return to page

Open-weight H2O-Danube3-4B

H2O.ai overtakes Apple and matches Microsoft with Danube3-4B, scoring over 80% accuracy on 10-shot HellaSwag benchmark

We're excited to announce that the H2O-Danube3 series is now available globally on Hugging Face. This latest series includes H2O-Danube3-4B and the compact H2O-Danube3-500M models.

The H2O-Danube3-4B, with its training on 6 trillion tokens, and the H2O-Danube3-500M, with 4 trillion tokens, are designed to handle extensive datasets and are fine-tuned for a multitude of applications. These models are crafted to bring advanced NLP capabilities to the masses by being efficient enough to operate on modern smartphones.

The H2O Danube3-4B model achieves an impressive score of over 80% on the 10-shot #HellaSwag benchmark, surpassing #AppleLLM OpenELM-3B-Instruct and competing with Microsoft Phi3 4B.

Additionally, the H2O-Danube3-500M model excels by scoring highest in 7 out of 12 academic benchmarks compared to models of similar size, such as Alibaba Qwen2-0.5B and Apple OpenELM-0.5B-Instruct.

These models exemplify extraordinary versatility and efficiency, making them perfect for a diverse array of applications including chatbots, research, and on-device solutions.

display of a laptop, desktop, tablet, cell phone, and IoT devices display of a laptop, desktop, tablet, cell phone, and IoT devices

Early H2O Danube2 applications

PII Detection

Detect patterns of personal identification Kaggle

LLM Generated Content Detection

Easier to detect human generated

Guardrails LLMs and Gateway LLM

LLM Safety with an LLM

Own Your Data

Post-train and fine-tune LLMs on your tokens for best price / performance on commodity hardware

H2O Danube-Powered Mobile App

H2O AI Personal GPT

Content Generation: Writing and editing in airplane mode.

Research: Analyzing and learning in offline mode. Accessing critical information while stranded.

Guardrails & Gateway: Confirm a user's question and input is valid and safe before sending to a more expensive model.

Entertainment: Reading pop culture trivia, learning historical facts, creating a social content calendar.

Remote Field Work (IoT): Technicians can get data from IoT sensors on their mobile devices in the field even during service blackouts.