H2O-Danube3

Danube3 is H2O.ai's third-generation small language model series, launched in July 2024 with a 4B-parameter and a 500M-parameter variant. The 4B model was trained on 6 trillion tokens (the 500M on 4 trillion), uses a Llama-2-style architecture with a Mistral tokenizer and an 8K context window, and was released under the Apache 2.0 license. H2O.ai reported over 80% on 10-shot HellaSwag for the 4B model, positioning it against Phi-3 and Apple's OpenELM-3B. The headline design goal is on-device inference — both models run natively and fully offline on a modern smartphone via H2O's own apps.\n\nThe honest picture: Danube3 is a general-purpose small LM, not a reasoning or instruction-tuned flagship, so expectations should match its size class. H2O.ai's benchmark framing compares against Apple's and Microsoft's smallest models of mid-2024, and the field has moved on — but as a fully permissive, phone-ready open model, Danube3 remains a practical choice for edge chat and offline NLP.