Reka has launched Reka Core, a new multimodal language model designed to compete with established models like GPT-4 and Claude 3 Opus. Reka Core, which started from scratch with no compute, no codebase, and no datasets, has been developed over several months and is now generally available. It boasts capabilities in understanding videos, advanced coding, and multilingual communication, matching or surpassing leading models in key evaluation metrics. Reka Core is comparable to GPT-4V, with a MMLU score of 83.2% versus GPT-4's 86.4%, and is currently undergoing further training. It outperforms Claude 3 Opus on multimodal human evaluations and surpasses Gemini Ultra on video tasks. Users can try Reka Core for free and learn more about its features through the company's technical report.