Hyperbolic Labs has announced the hosting of several advanced Vision Language Models (VLMs), including Pixtral-12B from MistralAI and Qwen2-VL-7B from Alibaba Qwen. These models are being served in BF16 and are particularly effective in handwriting recognition and OCR tasks, with Qwen2-VL-7B achieving 100% accuracy in tests. Qwen2-VL-7B also ranks 8th on the leaderboard for models under 10 billion parameters, making it the best open-source VLM in its category. Additionally, Qwen2-VL-Max-0809 (72B) outperforms GPT-4o-0806 and ranks 1st overall. Hyperbolic Labs has also retired Reflection 70B in favor of DeepSeek-V2.5, which combines DeepSeek-v2 and DeepSeek-v2-Coder, and is running on 8xH100 GPUs. The models are currently available for free through Hyperbolic Labs' API and playground. Qwen2-VL 70B has shown notable performance on DocVQA, OCRBench, and InfoQA.