Reka Unveils Multimodal Language Model Core with Industry-Leading Performance

AI startup Reka recently released its latest and cutting-edge multimodal language model - Core. This Core model, together with Reka's existing Flash model (with 21 billion parameters) and Edge model (with 7 billion parameters), forms Reka's model matrix. Reka refers to the Core model as a state-of-the-art model, demonstrating industry-leading performance in a wide range of tasks involving text, images, videos, and audio.

According to Reka's technical report, the Core model is primarily trained from scratch using PyTorch on NVIDIA H100s. It is worth noting that Reka states that the model is still undergoing continuous optimization and has not yet completed training. However, based on existing benchmark test results, the model performs on par with leading models from OpenAI, Anthropic, and Google, and even surpasses them in certain aspects. The model exhibits exceptional reasoning abilities, including complex language and mathematical capabilities, making it an ideal choice for complex analysis and problem-solving.

One notable feature of the Core model is its advanced multimodal understanding capabilities. Unlike many large-scale language models that primarily focus on text, the Core model has a deep and contextual understanding of images, videos, and audio. This makes it one of the only two commercial solutions on the market that provide comprehensive multimodal support.

In addition, the Core model boasts an impressive 128K context window, allowing it to ingest and accurately recall more information than many competitors. Combined with its outstanding language and mathematical reasoning abilities, this makes the Core model highly suitable for handling complex tasks that require in-depth analysis.

For developers, the Core model's top-notch code generation capabilities bring exciting possibilities for empowering autonomous workflows. The model's multilingual skills are equally impressive, as it can fluently use English as well as several Asian and European languages, thanks to its pretraining on text data from 32 languages.

Like Reka's other models, Core can meet the diverse needs of customers and partners through API, on-premises deployment, or device-side deployment. This flexibility, combined with Core's remarkable capabilities, unlocks a wide range of potential applications in industries such as e-commerce, social media, digital content, healthcare, and robotics.

Reka provides a webpage that includes multiple examples comparing the output of Core with GPT-4 and Claude 3 Opus. You can also explore the capabilities of these three models using their chatbot.