Reka Unveils Reka Core: Its Most capable Multimodal Language Model

Reka Unveils Reka Core: Its Most capable Multimodal Language Model

Reka, an AI startup based in Sunnyvale, California, has unveiled Core, its latest and most advanced multimodal language model developed to date. Their Core model joins their existing Flash (21B parameters) and Edge (7B parameters) models. Reka describes Core as a frontier-class model that delivers industry-leading performance across a wide range of tasks involving text, images, video, and audio.

According to Reka's technical report, Core was trained entirely from scratch primarily on NVIDIA H100s using Pytorch. Of note, Reka says the model is not finished training and is still improving. However, results from popular benchmarks so far, show that the model is on par with, or outperforms leading models from OpenAI, Anthropic, and Google. The model demonstrates exceptional reasoning skills, including complex language and math abilities, making it ideal for intricate analysis and problem-solving.

The model outperforms Claude-3 Opus in multimodal human evaluations conducted by an independent third party, surpasses Gemini Ultra on video tasks, and is comparable to GPT-4V on the challenging MMMU benchmark. On language tasks, Core is highly competitive with other frontier models across well-established benchmarks.

One of Core's standout features is its advanced multimodal understanding capabilities. Unlike many large language models that primarily focus on text, Core has a deep, contextualized understanding of images, videos, and audio. This makes it one of only two commercially available solutions that offer comprehensive multimodal support.

See more examples on Reka's website.

Core also boasts an impressive 128K context window, allowing it to ingest and accurately recall significantly more information than many of its competitors. Combined with its superb reasoning abilities across language and math, this makes Core well-suited for tackling complex, analytical tasks.

For developers, Core's top-tier code generation capabilities open up exciting possibilities for empowering agentic workflows. The model's multilingual skills are equally impressive, with fluency in English and several Asian and European languages, thanks to pretraining on textual data from 32 languages.

Like Reka's other models, Core can be deployed via API, on-premises, or on-device to meet the diverse needs of customers and partners. This adaptability, combined with Core's impressive capabilities, unlocks a wide range of potential use cases across industries such as e-commerce, social media, digital content, healthcare, and robotics.

Reka provided a page with various examples comparing Core's output to GPT-4 and Claude 3 Opus. You can also use their chatbot to explore the capabilities of their three models.

Let’s stay in touch. Get the latest AI news from Maginative in your inbox.