Reka
By Reka AI
Reka is a family of multimodal large language models developed by Reka AI, a startup founded by former DeepMind, Google Brain, and Meta AI researchers, offering models that process text, images, video, and audio inputs.
Definition
Reka is a family of multimodal large language models developed by Reka AI, a startup founded by former DeepMind, Google Brain, and Meta AI researchers, offering models that process text, images, video, and audio inputs.
Overview
Reka AI was founded in 2022 by researchers with backgrounds at DeepMind, Google Brain, and Meta AI (including Yi Tay and collaborators) to build frontier multimodal foundation models as an independent lab. The company released its Reka Core, Reka Flash, and Reka Edge model family in early 2024, spanning a range of sizes: Reka Core as its largest, most capable flagship model competitive with other frontier multimodal models of its era on public benchmarks, Reka Flash as a smaller, faster mid-tier model, and Reka Edge as a compact model intended for on-device or latency-sensitive deployment. A defining characteristic of the Reka model family is native multimodality: the models were trained to natively understand and reason over combinations of text, images, video, and audio inputs, rather than bolting vision capability onto a text-only base via a separate adapter. This let Reka models handle tasks such as answering questions about video content, interpreting charts and screenshots, and reasoning across mixed media inputs. Reka positioned its models for enterprise and API use cases, competing with the multimodal offerings of OpenAI, Google, and Anthropic, while emphasizing a smaller, independently-funded lab's ability to iterate quickly. Reka AI has continued to release updated model versions and has explored both proprietary API access and, for smaller models, more open licensing terms. As with many mid-sized foundation model labs, Reka competes in a crowded field dominated by hyperscaler-backed labs (OpenAI, Google DeepMind, Anthropic) and other well-funded independents (Mistral, Cohere), differentiating primarily on multimodal breadth, pricing, and focused enterprise deployment support.
Key Concepts
- Natively multimodal: trained jointly on text, image, video, and audio inputs
- Model family spans multiple sizes: Reka Core (flagship), Reka Flash (mid-tier), Reka Edge (compact)
- Founded by researchers from DeepMind, Google Brain, and Meta AI
- Positioned for enterprise API and applied multimodal reasoning use cases
- Competes directly with GPT-4o, Gemini, and Claude on multimodal benchmarks
- Supports video and chart/document understanding beyond static image inputs