Artificial Intelligence · Profile from AI Research Companion
Reka
A small multimodal model lab, founded in 2022 by researchers from Google DeepMind, Google Brain and Meta, that trains its own models to understand text, images, video and audio and sells them, and agent products built on them, to enterprises. Reka is the counter-example to the multi-billion-dollar "reputation round": it emerged from stealth with ~$58m (June 2023), turned down or lost a reported >$1bn acquisition by Snowflake (talks ended May 2024), lost its best-known co-founder back to Google DeepMind (November 2024), and became a unicorn only in July 2025 with a $110m round backed by Nvidia and Snowflake (~$1bn valuation, reported). With ~50 staff (July 2025, reported), it competes against labs hundreds of times larger by staying small, publishing efficient models (Reka Flash 3 with open weights, March 2025) and, from June 2026, turning towards "physical AI" after joining forces with a team from Moonvalley.
At a glance
- Founded
- 2022 by Dani Yogatama (CEO; previously at DeepMind), Yi Tay (chief scientist; previously at Google Brain), Qi Liu, Cyprien de Masson d'Autume and others from DeepMind, Google Brain and Meta (reported; verify full founder list)
- Headquarters
- San Francisco Bay Area (sources give San Francisco and Sunnyvale), with staff in Singapore, London and elsewhere (reported, verify)
- Ownership / structure
- Private. Investors include DST Global Partners and Radical Ventures (2023 leads), Snowflake Ventures, Nvidia (NVentures) and others (reported)
- Leadership (2026)
- Dani Yogatama (co-founder and CEO); Yi Tay left in November 2024 to return to Google DeepMind; Mateusz Malinowski and Mikołaj Bińkowski (ex-Google DeepMind, Veo contributors) joined via the Moonvalley team (June 2026)
- AI products
- Yasa-1 multimodal assistant (October 2023); Reka Flash (21bn parameters) and Reka Edge (7bn) (February 2024); Reka Core (April 2024); Vibe-Eval benchmark (May 2024); Reka Flash 3 open-weight reasoning model under Apache 2.0 (March 2025) and Flash 3.1 (2025); Reka Quant quantisation; Reka Research and Reka Vision agent platforms (generally available by mid-2025); "World Language Action Model" research for physical AI (2026)
- Scale
- ~20 staff growing to ~50 between mid-2024 and July 2025 (SiliconANGLE); revenue not disclosed; models distributed through its own API and partners including Snowflake Cortex (2024)
- Capital raised
- ~$58m at emergence from stealth (June 2023; Reuters reported $50m led by DST Global Partners); ~$300m valuation (PitchBook, 2023); $110m backed by Nvidia and Snowflake at ~$1bn (July 22, 2025, Bloomberg); total ~$170m (sum of reported rounds)
- Key partnerships
- Snowflake (investor, Cortex model host, acquisition suitor, supplier relationship); Nvidia (investor, compute); Moonvalley (team combination, June 2026); Shutterstock (image and video data licence, 2024, reported in Shutterstock's files; terms not disclosed, verify)
- Signature moves (incl. failures)
- Founding by big-lab researchers (2022); Series A and Yasa-1 (2023); Core, Flash and Edge (2024); the failed Snowflake acquisition (2024); co-founder Yi Tay's return to Google (2024); Flash 3 open weights and agent products (2025); the Nvidia–Snowflake $110m round (2025); the Moonvalley team and the physical-AI turn (2026)
7 strategic moves analysed
Each one gets a full strategy file in the companion: context, the decision, the financing, what happened, and the lesson.
2022
Founding by Big Lab Researchers
Founding
2023
Series A and Yasa 1
Funding round / launch
2024
Core Flash Edge Multimodal Models
Model release
2024
Snowflake Acquisition Talks Collapse
Acquisition (failed) / talent loss
2025
Flash 3 Open Weights and Agent Products
Model release / launch
2025
Nvidia Snowflake 110m Round
Funding round
2026
Moonvalley Team and Physical AI Turn
Team combination / pivot