H2O.ai’s Small Vision-Language Models Surpass 2.4 Million Monthly Downloads

H2OVL Mississippi models signal a shift from costly frontier APIs to self-hosted, higher-performing document AI.

MOUNTAIN VIEW, Calif., July 22, 2026 — H2O.ai, the world’s leading sovereign enterprise AI platform spanning predictive, generative and agentic AI, today announced that its open-weight H2OVL Mississippi vision-language models have surpassed a combined 2.4 million monthly downloads on Hugging Face. Purpose-built for optical character recognition and document AI, the two models are designed to run within the organization’s own infrastructure, and outperform models many times their size, a shift away from the costly, cloud-bound frontier LLMs that have dominated enterprise AI.

The pull is economic and reflects a broader shift in Enterprise AI. Reading documents with a general-purpose frontier model often means paying per token to send sensitive pages to a frontier LLM provider’s cloud. A purpose-built small model does the same work for a fraction of the cost, on the customer’s own infrastructure, with the data never leaving their walls. What makes the numbers notable is that neither model is new: the 800M model published in December 2024 and the 2B in 2025, and both still clear a million downloads every month. That points to sustained demand, not a launch-driven spike.

“The large model is becoming a commodity. The value is in small, purpose-built models you own and run where your data lives, not rented by the token from someone else’s cloud. Millions of downloads a month tell you where the market is moving. Download it, point it at your own documents, and watch your token bill fall. That’s why we put it in the open.”

Sri Ambati

CEO and Founder of H2O.ai

Why Small Models Are Winning Document AI

• Cost. Small models turn per-token API bills into fixed, owned infrastructure.
• Privacy. Documents stay inside the customer’s environment, which matters most in regulated industries.
• Deployability. At under two billion parameters, the models run where the data already lives, whether on-premises, in a private cloud, or air-gapped.
• Speed. Smaller, more focused LLM models can have shorter inference response times.

Inside H2OVL Mississippi

H2OVL Mississippi is a family of two open-weight vision-language models, both released under the Apache 2.0 license and built on H2O.ai’s H2O Danube language models.

H2OVL-Mississippi-2B is a high-performing, general-purpose vision-language model with 2 billion parameters, built for image captioning, visual question answering, and document understanding while staying efficient enough for real-world deployment. Trained on 17 million image-text pairs, it ranks competitively among 2B-class models, leading its size class on MathVista (56.8) and scoring 782 on OCRBench, within a point of the larger Qwen2-VL-2B (797).

H2OVL-Mississippi-800M is a compact 0.8-billion-parameter model trained on 19 million image-text pairs focused on OCR, document comprehension, and chart, figure, and table interpretation. On the text-recognition segment of OCRBench, it scores 274 out of 300, ahead of models many times its size, including the 26-billion-parameter InternVL2-26B.

Text Recognition Results on OCRBench — H2OVL Mississippi-0.8B (274) tops all SLMs and larger LLMs

H2OVL Mississippi-0.8B leads on OCRBench text recognition, beating models up to 30x its size. Source: H2O.ai. Benchmark set as of model release; 274/300 across six text-recognition categories.

Open on Hugging Face, Production-Grade in the Enterprise

Hugging Face is where practitioners discover, test, and benchmark models. H2O.ai is where regulated enterprises bring them into production, with the governance, deployment, and support that mission-critical document workloads require. The efficiency is not theoretical: H2O.ai helped a large North American telecom operator reduce token costs by 10x by moving from a frontier model approach to a small language model fine-tuned on the company’s own data and hosted on its own hardware.

The download numbers reflect a broader shift: as the cost of frontier tokens meets the reality of enterprise-scale document processing, purpose-built small models are becoming the default for the work that does not need a giant model.

Benchmark fact sheet

H2OVL-Mississippi-2B — OpenVLM benchmark comparison (2B class)

H2OVL-Mississippi-800M — OCRBench text recognition 274/300 · composite OCRBench 751 · trained on 19 million image-text pairs.

Model collection:

https://huggingface.co/collections/h2oai/h2ovl-mississippi-66e492da45da0a1b7ea7cf39