Blogs›H2O.ai›H2O Danube Large Language Models
H2O LLM Studio provides a no-code framework for fine-tuning state-of-the-art Large Language Models (LLMs) and Small Language Models (SLMs). It leverages Deepspeed for distributed training on GPU clusters, enabling cost-effective and faster training compared to traditional LLMs. The studio supports various fine-tuning techniques including instruction/chat fine-tuning, causal classification and regression, and DPO/IPO/KTO optimization. It also facilitates distilling LLMs into SLMs using H2O LLM Data Studio, leading to lower TCO and potentially higher accuracy for specific use cases.
6/24/2026
This post introduces H2O LLM Studio, a no-code fine-tuning framework for creating custom LLMs and SLMs. It details the training stages and data mixes used for H2O-Danube3-4B, highlighting the benefits of SLMs (cheaper, faster, more customizable) over LLMs. It also outlines various fine-tuning techniques supported by the studio, such as instruction/chat fine-tuning, causal classification/regression, and DPO/IPO/KTO optimization.
6/24/2026
This post introduces h2oGPTe Agentic AI, an extension of H2O's LLM capabilities. It details the integration of Generative AI with Predictive AI through a multi-agent platform. Key technical contributions include multimodal analysis (audio/vision), a coding assistant for rapid prototyping, autonomous agentic AI for multi-step workflows, citation-based Retrieval Augmented Generation (RAG) for transparency, customizable guardrails for AI safety, intelligent model routing for optimal LLM selection, and model risk management with embedding-based metrics and human feedback calibration. It also highlights on-premise and air-gapped deployment options, scalability via Kubernetes, and performance metrics like 100+ queries/minute on single GPU deployments.
6/24/2026
This post introduces the H2O Danube3 series of open-weight large language models. It details the training process involving three stages with varying data mixes and token counts (4.6T, 1.35T, and 0.05T tokens). The models are trained on approximately 6 trillion tokens using ~100 H100 GPUs and H2O LLM Studio. Performance highlights include Danube3-4B scoring over 80% on the 10-shot HellaSwag benchmark, surpassing AppleLLM OpenELM-3B-Instruct and competing with Microsoft Phi3 4B. Danube3-.5B outperforms Alibaba Qwen2-.5B and Apple OpenELM-.5B Instruct in 7 out of 12 academic benchmarks. The post also discusses applications such as cost-efficient on-device processing, enhanced privacy through local data processing, AI content detection, and guardrail LLMs for GenAI safety. The H2O AI Personal GPT mobile app is presented as an example application.