Reka Core

À propos de Reka AI

Fondée 2022 · San Francisco, California, USA

Reka AI was founded in mid-2022 by Dani Yogatama (CEO, ex-Google DeepMind, ex-Facebook), Yi Tay (Chief Scientist, ex-Google Brain), Donovan Ong (ex-Apple) and Qi Liu, with engineering teams in San Francisco, London and Singapore. The founders had previously worked on large-scale language models, mixture-of-experts and efficient long-context architectures inside DeepMind and Brain. Reka raised a $58M Series A in 2023 led by DST Global Partners and reportedly closed a Series B at a near-unicorn valuation in 2024. The Reka model line is intentionally multimodal-first and includes Edge (small), Flash (mid) and Core (flagship), all trained jointly on text, image, video and audio. Reka Core launched in April 2024 with the public Reka Core technical report, positioning it as the first multimodal-from-the-ground-up frontier model from a startup outside the FAANG/DeepMind orbit.

Architecture

Decoder-only Transformer trained multimodally on text, image, video and audio

Reka Core is the flagship of Reka's multimodal frontier family. According to the publicly released Reka Core technical report, the model is a decoder-only Transformer trained from scratch on a multimodal corpus comprising text, code, images, video frames and audio clips, with separate modality encoders that project into a shared token embedding space. Training was carried out on a custom cluster using Pathways-style infrastructure and a curriculum that first pretrains on text and code, then progressively introduces images, audio and video. The context window is 128K tokens, video input is supported up to several minutes, and audio is processed via a learned audio encoder (described as ImageBind-style on the audio side). The training compute is reported as roughly an order of magnitude less than GPT-4 while reaching competitive scores on MMMU, Perception Test and VideoMME. Core is offered exclusively through the Reka API, Reka Playground and Snowflake Cortex; weights are not released.

Paramètres

Undisclosed (described as 'flagship', estimated tens of billions)

Contexte

128 000 tokens

Capacités

• True multimodal input: text, image, video and audio in the same context
• 128K-token context window
• Video understanding up to several minutes
• Audio understanding for transcription and sound classification
• Function calling, JSON output and tool use
• Available on Reka API, Reka Playground and Snowflake Cortex
• Multilingual coverage across 32+ languages
• Best for: multimodal assistants, video QA, audio-grounded chat, frontier features at startup pricing

Entraînement et licence

Decoder-only multimodal training over a curated corpus of text, code, images, video frames and audio with progressive curriculum. Compute reported as roughly an order of magnitude less than GPT-4.

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

Tests de sécurité: The Reka Core technical report includes safety, bias and toxicity evaluations; the API integrates safety filters.