Reka Edge

Informazioni su Reka AI

Fondato 2022 · San Francisco, California, USA

Reka AI was founded in mid-2022 by Dani Yogatama, Yi Tay, Donovan Ong and Qi Liu, all senior researchers from Google DeepMind, Google Brain, Apple and Facebook AI. The company is headquartered in San Francisco with engineering centres in London and Singapore. Reka has raised $58M in Series A funding led by DST Global Partners in 2023 and closed an additional Series B in 2024 at a reported near-unicorn valuation. The Reka model family is multimodal-by-design (text, image, video, audio) and is sized into three tiers: Edge (smallest), Flash (mid), Core (flagship). Reka Edge (7B parameters) was released alongside Flash and Core in early 2024 as the on-device / cost-sensitive sibling, designed to run on a single consumer GPU while preserving the multimodal capabilities that distinguish Reka from text-only competitors.

Architettura

Decoder-only Transformer trained multimodally on text, image, video and audio (small variant)

Reka Edge is a ~7B parameter decoder-only Transformer trained on the same multimodal corpus as the larger Reka Flash and Core, comprising text, code, images, video frames and audio clips. According to the public Reka Core technical report, all three sizes share the architecture (modality encoders projecting into a shared token embedding space) but differ in scale and the proportion of training compute spent on each modality. Edge is optimised for on-device and cost-sensitive deployments, fits on a single 24 GB consumer GPU in FP16 and supports a 128K-token context window. It accepts text, image, video (up to short clips) and audio (with a learned audio encoder). On MMMU, Perception Test and VideoMME, Edge scores below Flash and Core but ahead of similarly sized open-weights multimodal baselines like LLaVA-Next 7B and InternVL 7B at launch. Edge is offered through the Reka API and Reka Playground; weights are not publicly released.

Parametri

7B

Contesto

128.000 token

Capacità

• Multimodal input: text, image, short video and audio
• 128K-token context window
• Runs on a single 24 GB consumer GPU in FP16
• JSON output and tool use
• Multilingual coverage across 32+ languages
• Hosted through Reka API and Reka Playground
• Cost-efficient pricing tier in the Reka family
• Best for: cost-sensitive multimodal apps, on-device experiments, edge AI

Addestramento e licenza

Multimodal pretraining over a curated corpus of text, code, images, video frames and audio with progressive curriculum, identical mix to Reka Core but smaller compute budget.

Licenza: Proprietary commercial API. Generated outputs may be used commercially under the Reka terms.

Test di sicurezza: Same safety and bias evaluation framework as the Reka Core technical report; safety filters integrated in the Reka API.