Reka AI

  Executive Briefing

Reka AI is an independent multimodal AI research and product company founded in August 2022 by five researchers who collectively assembled one of the most credentialed founding teams in the post-GPT-3 startup wave. The company set out to do something no major lab had done at small scale: build foundation models that process text, images, video, and audio natively from the ground up, rather than retrofitting vision or audio onto a language-first backbone. Headquartered in Sunnyvale, California, Reka is led by co-founder and CEO Dani Yogatama (ex-Senior Staff Research Scientist, Google DeepMind) and co-founder and CTO Cyprien de Masson d'Autume (ex-DeepMind, Gopher and AlphaCode), supported by co-founders Mikel Artetxe (ex-Meta AI FAIR, multilingual NLP specialist) and Qi Liu (ex-Google DeepMind, Meta AI FAIR, and Microsoft Research).¹

In under two years, Reka's team of roughly twenty researchers built Reka Core — a frontier-class multimodal model that ranked second in blind human evaluation of multimodal chat and outperformed Gemini Ultra on video question-answering — and followed it with the open-source Reka Flash 3, a 21-billion-parameter reasoning model competitive with OpenAI o1-mini, released under Apache 2.0 in March 2025.² The company reached unicorn status in July 2025 when it closed a reported $110M Series B led by NVIDIA and Snowflake at a reported $1 billion valuation.³

June 2026 marked a strategic pivot: Reka merged with Toronto-based video-AI startup Moonvalley, whose co-founders Mateusz Malinowski and Mikołaj Bińkowski — both former Google DeepMind staff researchers who helped build the predecessor to Google Veo 2 — joined Reka's leadership.⁴ The combined entity is now building the World Language Action Model (WLAM), an omni-model trained on egocentric and physical-world data targeting robotics, autonomous systems, defense, and intelligent wearables. Reka has restructured around four pillars: Reka Labs (foundational research), Reka Infer (inference API), Reka Vision (enterprise video/image platform), and Reka Claru (physical-AI training data).⁴

  At a Glance

  Origins & Founding

Reka was incorporated in August 2022 in Sunnyvale, California, by five researchers who had collectively shaped some of the most influential AI projects of the preceding decade. The founding thesis was precise: general-purpose large language models as deployed by leading labs were impractical for the full range of enterprise use cases, and the market opportunity lay in building truly multimodal, computationally efficient foundation models that enterprises could customize on their own proprietary datasets without sacrificing privacy.¹

Dani Yogatama, who led the effort as CEO, had spent six years at Google DeepMind (2016–2022) contributing to AlphaStar and DeepSpeech and had earlier worked at Baidu's Silicon Valley AI Lab; he holds a PhD from Carnegie Mellon University and concurrently serves as Associate Professor of Computer Science at USC.⁵ Cyprien de Masson d'Autume (CTO) was a DeepMind Staff Researcher (2016–2022) with direct experience on Gopher and AlphaCode. Yi Tay, the founding Chief Scientist, had co-led PaLM-2 architecture, UL2, Flan-2, and the Differentiable Search Index at Google Brain. Mikel Artetxe, a specialist in multilingual NLP and unsupervised machine translation, came from Meta AI FAIR (2020–2023) and holds a PhD from the University of the Basque Country, where he remains an honorary researcher.⁶ Qi Liu, the fifth co-founder, had worked across Google DeepMind, Meta AI FAIR, and Microsoft Research, and holds a PhD from the University of Oxford; he is also Assistant Professor of Computer Science at the University of Hong Kong.

The founding mission was to build "complete multimodal AI solutions" for enterprises, anchored in four research pillars: universal intelligence, general-purpose multimodal and multilingual agents, self-improving AI, and model efficiency. Reka's defining technical bet — training natively multimodal models that ingest text, image, video, and audio from the first token rather than bolting on modalities later — was unusual for an early-stage startup and placed it in direct philosophical opposition to the adapter-and-fine-tune approach common among enterprise AI vendors.

  History & Timeline

    2022–2023: Stealth, seed, and Series A

Reka operated in stealth for nearly a year after incorporation, building its initial research infrastructure and multimodal assistant. In June 2023, the company emerged from stealth announcing a $58M Series A led by DST Global Partners and Radical Ventures, with Snowflake Ventures and angel investor Nat Friedman (former CEO of GitHub) participating.¹ At the time of the announcement, Yasa, Reka's multimodal assistant, was in closed beta. An additional reported tranche of ~$50M from DST Global Partners and Snowflake followed, bringing total 2023 funding to approximately $108M at a valuation of around $300M, though the exact split between tranches is reported differently across sources.⁷

By October 2023, Reka publicly launched Yasa-1: a multimodal assistant supporting 20 languages, 100,000-token context, image/video/audio input, and in-session code execution — and benchmarked at up to 8x faster than Claude 2 on comparable tasks. Distribution was available via API and Docker for on-premises deployment.¹

    2024: Core, Flash, and Edge — the model family

April 2024 was Reka's most active publication month. The company announced a compute partnership with Oracle, selecting Oracle Cloud Infrastructure's OCI Supercluster (up to 32,768 H100 GPUs) as its primary training environment, with Reka Core and Flash listed on the Oracle Cloud Marketplace.⁸ Simultaneously, Reka published arXiv tech report 2404.12387, a 25-author paper introducing the Core, Flash, and Edge model family: Core benchmarked second in blind human multimodal chat evaluation and outperformed Gemini Ultra on Perception-Test video QA; Flash was a 21B-parameter mid-tier model; Edge was a 7B-parameter model designed for on-device and resource-constrained deployment.²

May 2024 brought two significant developments. Reka released Vibe-Eval — an open visual evaluation benchmark of 269 prompts (100 hard-difficulty) with expert-authored gold responses — on GitHub.⁹ That same month, Bloomberg reported that Snowflake was in acquisition talks to acquire Reka for $1 billion or more, representing a 3x+ premium on the then-current valuation. The talks collapsed within a week, on 22 May 2024, with both parties remaining independent — an episode that demonstrated both Reka's perceived strategic value and the complications of a potential acqui-hire at that scale.¹⁰

In November 2024, co-founder and Chief Scientist Yi Tay departed Reka to return to Google DeepMind as Senior Staff Research Scientist, where he subsequently co-led the Gemini Deep Think model (which achieved a gold medal at the International Mathematical Olympiad). Tay's departure removed one of the core architects of Reka's model quality strategy.¹¹

    2025: Open-source, unicorn status, and product expansion

March 2025 brought Reka's highest-profile open-source release: Reka Flash 3, a 21B-parameter reasoning model trained with RLOO (Reinforce Leave One Out) reinforcement learning, released under the Apache 2.0 license with weights hosted on Hugging Face under the RekaAI organization.³ Flash 3 benchmarked competitively with OpenAI o1-mini on reasoning tasks.

In July 2025, Reka closed its $110M Series B led by NVIDIA and Snowflake at a reported $1 billion valuation — a roughly 3.3x step-up from the ~$300M valuation of two years prior, and a milestone that made Reka one of a small number of AI unicorns built by sub-100-person teams.³ Reka Flash 3.1 was released concurrently, along with Reka Quant — a quantization technology Reka claimed reduced model size with only 1.6 points of benchmark degradation versus 6.7 points for comparable alternatives (claim sourced from Reka's own reported figures; independently unverified).³ Reka Vision and Reka Research entered general availability, with Shutterstock and Turing Video named as reference customers.

    2026: Moonvalley merger and the physical-AI pivot

On 11 June 2026, Reka and Toronto-based Moonvalley announced a merger via all-share exchange (financial terms not disclosed).⁴ Moonvalley co-founders Mateusz Malinowski (PhD, former Google DeepMind Staff Research Scientist, co-creator of a predecessor to Google Veo 2) and Mikołaj Bińkowski (PhD, former Google DeepMind Senior Research Scientist) joined Reka's founding team. The merger was framed as a strategic combination: Reka's multimodal model and inference stack with Moonvalley's video generation and physical-world modeling expertise. The combined entity announced the World Language Action Model (WLAM) — an omni-model trained on egocentric and physical-world data — and restructured into four pillars: Reka Labs, Reka Infer, Reka Vision, and Reka Claru.⁴

  Mission, Philosophy & Research Agenda

Reka's stated mission is to "advance science and build generative AI models for the benefit of humanity, organizations, and enterprises." The founding philosophy centers on four research pillars: universal intelligence, general-purpose multimodal and multilingual agents, self-improving AI, and model efficiency. Reka's approach is defined by a conviction that genuine multimodality — training models to perceive and reason over text, image, video, and audio simultaneously from the ground up — produces qualitatively different and more capable systems than single-modality models with bolted-on adapters.

Efficiency is equally central to the research agenda: Reka has consistently prioritized training and inference efficiency, aiming to match frontier performance at lower parameter counts and cost. This is reflected in the Core/Flash/Edge tiering, the RLOO reinforcement-learning post-training for Flash 3, and the Reka Quant quantization work. The post-2026 merger has expanded the mission to encompass "physical AI" — models and infrastructure that reason, simulate, and act in the physical world — targeting robotics, autonomous systems, defense, and next-generation media.

  Models & Products

Reka's model family progresses across three tiers, with open-source releases at the mid-tier beginning in 2025. The product surface spans developer API access, an enterprise video/image understanding platform, an agentic research tool, and a creator-focused video-clipping tool.

Model family:

• Reka Core — Largest and most capable model (announced April 2024); benchmarked second in blind human multimodal chat evaluation; competitive with GPT-4V on image QA; outperforms Gemini Ultra on Perception-Test video QA.²
• Reka Flash — 21B-parameter mid-tier multimodal model (announced April 2024); processes text, images, video, and audio natively.
• Reka Edge — 7B-parameter efficiency-focused multimodal model designed for on-device and resource-constrained deployment.
• Reka Flash 3 — Open-source 21B reasoning model (March 2025); Apache 2.0 license; trained with RLOO reinforcement learning; competitive with OpenAI o1-mini; weights on Hugging Face.³
• Reka Flash 3.1 — Updated reasoning model (July 2025); reported to benchmark competitively against Qwen3-32B and OpenAI o3-mini at lower cost.³
• Reka Quant — Quantization technology for production model compression (July 2025).
• World Language Action Model (WLAM) — Announced June 2026; omni-model trained on egocentric and physical-world data; no release date, parameter count, or benchmark results disclosed at time of writing.⁴

Products and platforms:

• Reka Vision — Enterprise platform for video and image processing: natural language search, video summarization, and real-time event detection; customers include Shutterstock (content library metadata enrichment) and Turing Video (Guardian AI agentic surveillance for US law enforcement).³
• Reka Research — Agentic AI for complex question-answering; browses the web and private documents (including Google Drive) and cites sources transparently.
• Reka for Creators — Automated tool converting long-form video content into social media clips with captions and hashtags.
• Reka Infer — Post-merger inference infrastructure and API division.
• Reka Claru — Post-merger training data division for physical AI (egocentric video, robotics trajectories, world-model footage, expert human judgment).
• Yasa-1 — Original multimodal assistant product (October 2023); supported 20 languages, 100K-token context, image/video/audio input, and code execution; available via API and Docker. Superseded by the Core/Flash/Edge family.

  People

Reka's founding team is unusually research-dense: all five co-founders hold PhDs and carried direct experience from frontier model projects at Google DeepMind, Google Brain, and Meta AI FAIR before launching the company. The current leadership combines the four remaining original co-founders with two more from the Moonvalley merger.

Co-founders and current leadership:

• Dani Yogatama (CEO) — ex-Senior Staff Research Scientist at Google DeepMind (2016–2022, AlphaStar, DeepSpeech) and Baidu Silicon Valley AI Lab; PhD Carnegie Mellon; also Associate Professor of Computer Science at USC.⁵
• Cyprien de Masson d'Autume (CTO) — ex-DeepMind Staff Researcher (2016–2022, Gopher, AlphaCode).
• Mikel Artetxe (Co-founder) — ex-Research Scientist at Meta AI FAIR (2020–2023); PhD University of the Basque Country; specialist in multilingual NLP and unsupervised machine translation; Honorary Researcher at the University of the Basque Country (IXA group).⁶
• Qi Liu (Co-founder) — ex-Google DeepMind, Meta AI FAIR, Microsoft Research; PhD University of Oxford; also Assistant Professor of Computer Science at the University of Hong Kong.
• Mateusz Malinowski (Co-founder, joined via Moonvalley merger, June 2026) — PhD; former Google DeepMind Staff Research Scientist; co-creator of a predecessor to Google Veo 2.⁴
• Mikołaj Bińkowski (Co-founder, joined via Moonvalley merger, June 2026) — PhD; former Google DeepMind Senior Research Scientist.⁴

Notable departure:

• Yi Tay (Co-founder and initial Chief Scientist) — departed November 2024 to return to Google DeepMind as Senior Staff Research Scientist; subsequently co-led the Gemini Deep Think model (IMO gold medal, 2025) in Singapore's GenAI research lab.¹¹ Tay's departure was a significant loss, as he had been central to the design of Reka's model architecture and quality.

Headcount is estimated at 50–60 as of late 2025 / early 2026, a figure that is fast-moving post-Moonvalley merger. Revenue per employee is estimated at approximately $181K based on reported ARR figures, though these estimates are from third-party sources and are not independently verified.⁷

  Funding, Ownership & Business

Reka has raised a reported total of approximately $168M across two formal funding rounds, though some sources report $170M depending on how the two tranches of the 2023 round are counted.¹³⁷

The Series B valuation of $1 billion is widely reported but was not disclosed explicitly in Reka's own press release — it should be treated as reported rather than confirmed. NVIDIA serves the dual role of strategic investor and GPU infrastructure partner (H200/B200 inference optimization). Snowflake is a strategic investor and distribution partner, with Reka models accessible within the Snowflake platform.

Reka's business model combines token-based model-as-a-service (MaaS) through its API, enterprise platform licensing (Reka Vision, Reka Research, Reka Claru), and strategic distribution through the Oracle Cloud Marketplace and cloud hyperscaler channels. Reported ARR as of September 2025 stands at $10.9M (per Latka; self-reported or estimated, not independently verified).⁷ The Moonvalley merger introduces an all-share structure with no cash consideration publicly disclosed.

  Partnerships & Ecosystem

Reka's commercial strategy is built on a triangulation of compute infrastructure, cloud distribution, and enterprise customer deployment. The most significant infrastructure relationships are with Oracle and NVIDIA: Oracle's OCI Supercluster (announced April 2024) provides up to 32,768 H100 GPUs for training at a scale rare for a startup of Reka's size, and Reka Core and Flash are listed on the Oracle Cloud Marketplace for enterprise distribution.⁸ NVIDIA functions as both investor and optimization partner for inference on H200 and B200 hardware.³

Snowflake occupies a unique position as both a strategic investor (Series A and B) and a distribution channel, making Reka's models accessible to Snowflake's enterprise customer base — the same company that explored a full acquisition of Reka in May 2024. Enterprise customer deployments include Shutterstock, which uses Reka Vision to enrich and tag its content library, and Turing Video, which has built Guardian AI — an agentic video-surveillance product for US law enforcement agencies — on top of Reka Vision.³

  Notable Events & Controversies

Two events stand out in Reka's short history as significant inflection points. The May 2024 Snowflake acquisition talks — reported by Bloomberg at a price of $1 billion or more, roughly 3x Reka's prior valuation — collapsed within a week.¹⁰ The episode was notable as evidence that Snowflake viewed Reka as strategically indispensable but that the two parties could not reach agreement; it accelerated Reka's pursuit of independent Series B financing that ultimately closed over a year later at the same target valuation.

The November 2024 departure of Yi Tay — co-founder and Chief Scientist — was the more consequential internal event. Tay's work on PaLM-2, UL2, and Flan-2 at Google Brain had made him one of the most cited researchers in large-model architecture, and his subsequent achievement leading Gemini Deep Think at DeepMind underscored what Reka had lost.¹¹ More broadly, the June 2026 Moonvalley merger has been noted by industry observers as symptomatic of consolidation pressure among sub-frontier model startups: building and maintaining foundation models is capital-intensive, and compute access is increasingly difficult for labs that cannot achieve hyperscaler-level economics.

  Outlook & Roadmap

Reka's near-term trajectory is defined by the Moonvalley merger and the physical-AI pivot. The flagship research program is the World Language Action Model (WLAM) — an omni-model trained on egocentric and physical-world data for perception, action, and simulation — though no release date, parameter count, or benchmark results have been disclosed as of June 2026.⁴ The Reka Claru data platform, focused on egocentric video, robotics trajectories, and expert human judgment annotations, signals an intent to build a proprietary training data moat that competitors relying solely on internet-scraped text and image data cannot easily replicate.

The four-pillar organizational structure (Labs, Infer, Vision, Claru) indicates a platform strategy rather than a single-product bet: foundational research feeding into a productized inference API, a vertical enterprise platform for video understanding, and a data flywheel for physical AI. With $110M of Series B capital reported as freshly deployed, the company is scaling enterprise adoption of Reka Vision and Research into new verticals — including defense, public safety, and autonomous systems — while the Moonvalley team advances robotics and world-model capabilities. The addition of Malinowski and Bińkowski, who have direct experience with large-scale video generation at Google DeepMind, substantially strengthens Reka's technical bench for this direction.

Whether Reka can sustain the pace of its first three years — unicorn status, frontier-class models, an open-source release, and a merger, all with fewer than 60 employees — will depend on its ability to convert enterprise ARR into the compute budget required for physical-AI-scale training runs, and to attract the specialized talent (robotics, embodied AI, world simulation) its new mission demands. The consolidation pressure that drove the Moonvalley merger is real: the next phase of Reka's story will test whether a lean, research-first team can compete in a domain where Google DeepMind, Meta, and well-funded robotics startups are simultaneously building.

---

  References

1. Reka emerges from stealth — TechCrunch, Jun 2023
2. Reka Core, Flash, and Edge — arXiv 2404.12387
3. Reka AI raises $110M at $1B valuation — SiliconAngle, Jul 2025
4. Reka and Moonvalley merge for physical AI — PR Newswire, Jun 2026
5. Dani Yogatama — personal site
6. Mikel Artetxe — personal site
7. Reka AI financial data — Latka
8. Oracle and Reka collaboration — PR Newswire, Apr 2024
9. Vibe-Eval benchmark — GitHub
10. Snowflake acquisition talks — SiliconAngle, May 2024
11. Yi Tay — about page
12. Reka Series B announcement — Reka press release
13. Reka emerges from stealth — The Decoder

  References

1. Reka founding story, Series A announcement, and Yasa-1 launch — TechCrunch, Jun 2023. ↩ ↩² ↩³ ↩⁴ ↩⁵

2. Reka Core, Flash, and Edge model family — arXiv 2404.12387. ↩ ↩² ↩³

3. Series B round, valuation, Flash 3.1, Reka Quant, and customer details — SiliconAngle Jul 2025 and Reka press release. Valuation is reported; not independently confirmed in Reka's own press release. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴

4. Reka–Moonvalley merger and physical-AI pivot — PR Newswire, Jun 2026. WLAM details are as announced; no release date, parameter count, or benchmarks disclosed at time of writing. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

5. Dani Yogatama biography — dyogatama.github.io. ↩ ↩²

6. Mikel Artetxe biography — mikelartetxe.com. ↩ ↩²

7. Total funding figure varies across sources ($168M–$170M) depending on how the two 2023 tranches are counted; ARR of $10.9M as of September 2025 is per Latka (self-reported or estimated, not independently verified) — Latka. ↩ ↩² ↩³ ↩⁴

8. Oracle–Reka compute partnership — PR Newswire, Apr 2024. ↩ ↩²

9. Vibe-Eval open benchmark — github.com/reka-ai/reka-vibe-eval. ↩

10. Snowflake–Reka acquisition talks and breakdown — SiliconAngle, May 2024. ↩ ↩²

11. Yi Tay departure and subsequent work at Google DeepMind — yitay.net/about. ↩ ↩² ↩³