Reka AI

Reka AI (commonly referred to as Reka) is an artificial intelligence research and product company founded in 2022 by former Google DeepMind, Google Brain, Meta FAIR, and Baidu researchers.^[2][17][35] It began by building natively multimodal large language models that process text, images, video, and audio in a single model, and since 2026 has described itself as "an AI research lab building foundational intelligence for the physical world."^[1][29] Led by co-founder and CEO Dani Yogatama, Reka is best known for the Reka Core, Reka Flash, and Reka Edge model family and for the open-weights Reka Flash 3, and it became a unicorn in July 2025, when Bloomberg reported that a $110 million round backed by Nvidia and Snowflake valued it at more than $1 billion.^[3][31] Bloomberg reported in May 2024 that Snowflake was in talks to acquire Reka for more than $1 billion; the talks ended without a deal.^[5][31][32]

Reka emerged from stealth in June 2023 with $58 million in funding, according to TechCrunch, and released a family of models in different size classes: Reka Core, Reka Flash, and Reka Edge.^[2] It later added products such as Reka Nexus (an AI workforce platform), Reka Vision (video and image understanding), and Reka Research (an agentic web research tool).^[18][19][42] In mid-2026 it combined with the video-generation startup Moonvalley, in what Reka called a merger that added a team of researchers and engineers and The Information reported as an all-stock acquisition, and it turned toward world models, robotics, and data for physical AI.^[29][49] In October 2026 it released a research preview of Rho-1, a 19-billion-parameter model that both understands and generates text, images, and video and outputs robot actions.^[30]

What is Reka AI?

Reka AI is an AI lab that trains its own multimodal foundation models from scratch, meaning a single model is trained on text, images, video, and audio together rather than adding separate encoders to a text-only base.^[1] The company has stressed efficiency (its models were built by a small team that, according to SiliconANGLE, grew from about 20 to about 50 employees in the year before July 2025), flexible deployment (API, on-premises, and on-device), and open releases such as Reka Flash 3 under the Apache 2.0 license.^[10][16][31]

As of October 2026, Reka describes four lines of business: Reka Labs (research), Reka Infer (inference infrastructure and API for multimodal AI), Reka Vision (video processing infrastructure to tag, reason over, search, and clip large volumes of video), and Claru (training data for physical AI, including egocentric video, robot trajectories, world-model footage, and expert human judgment).^[29] Its homepage also promotes a Reka Cloud service to train, customize, and deploy models, and describes the company as building "models and infrastructure for the physical AI era" and "omni models that natively process and generate every kind of data in one architecture."^[60]

History

Founding and early development

Reka was founded in 2022 and incorporated in Delaware; its first SEC Form D filing reports a first sale of equity on September 29, 2022.^[35] TechCrunch named four co-founders in June 2023: Dani Yogatama, Cyprien de Masson d'Autume, Qi Liu, and Yi Tay.^[2] Mikel Artetxe, a former research scientist at Meta's FAIR lab, also lists himself as a Reka co-founder from 2023.^[13] According to TechCrunch, the co-founders said that while working on AI systems including DeepMind's AlphaCode and Google's Bard, they had concluded that it was impractical to expect one large language model to be deployed for every possible use case, such as writing ad copy in a particular brand's style.^[2]

Emergence from stealth (June 2023)

On June 27, 2023, Reka emerged from stealth. TechCrunch reported $58 million in funding led by DST Global Partners and Radical Ventures, with participation from strategic partner Snowflake Ventures and angel investors including former GitHub CEO Nat Friedman.^[2] Reka's own announcement gave the figure as $60 million, led by DST Global Partners and "founding investor" Radical Ventures.^[33] Two Form D filings made with the SEC in July 2023 report equity sales of about $7.45 million (first sale September 2022) and about $49.8 million (first sale June 2023).^[34][35] Bloomberg later reported that the 2023 financing valued Reka at about $300 million.^[31]

At the time, Reka's first product, Yasa, was in closed beta, and the company also offered to adapt its models to customers' own data and let them run the "distilled" models on their own infrastructure or through Reka's API.^[2][33] Yogatama told TechCrunch that Reka was not yet generating revenue and would spend its funding on computing power from Nvidia and on building a business team.^[2] Snowflake was an early customer as well as an investor, and Appen had announced it was working with Reka on multimodal applications.^[2]

Yasa-1 launch (October 2023)

In October 2023, Reka launched Yasa-1, its first publicly available multimodal AI assistant.^[9] Yasa-1 could understand images, short videos, and audio snippets as well as text, and supported 20 languages.^[9] The Decoder reported that it could process long documents of up to 100,000 tokens, with 24,000 tokens as the default context window, and could draw on web search results and execute code.^[37]

Yasa-1 was made available through APIs and as Docker containers for on-premise or VPC deployment, and it could be customized on private datasets of any modality.^[9]

Reka Flash and Reka Edge release (February 2024)

On February 12, 2024, Reka introduced Reka Flash, a 21-billion-parameter multimodal language model trained from scratch that it called the "turbo-class" offering in its lineup, together with Reka Edge, a smaller 7-billion-parameter model for resource-constrained settings such as on-device and local use.^[36] Both were released in public beta in the Reka Playground. Reka said Flash was competitive with Gemini Pro and GPT-3.5 on language and vision benchmarks, and that both chat models had been instruction tuned and then trained with reinforcement learning from human feedback using PPO, with Reka Flash serving as the reward model.^[36]

Snowflake Cortex partnership (March 2024)

On March 21, 2024, Snowflake announced that it was expanding its partnership with Reka to offer Reka's multimodal models in its Snowflake Cortex service, starting with Flash and with support for Core in development.^[15] Snowflake's AI product head told VentureBeat that more than 400 enterprises were using Cortex and the models it hosted, and that Reka's models would bring the number of LLMs on the service to a dozen.^[4]

Reka Core, Flash, and Edge technical report (April 2024)

On April 15, 2024, Reka announced Reka Core, its largest model, and published a technical report detailing the full lineup.^[1][16] The report, titled "Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models," was released as a preprint on arXiv.^[1]

According to the report, Reka Core approached the performance of GPT-4V on image question-answering benchmarks such as MMMU and VQAv2, and it ranked as the second most preferred model (behind GPT-4V) in a blind third-party human evaluation of multimodal chat, ahead of Claude 3 Opus.^[1] On video question answering (Perception-Test), both Core and Flash outperformed Gemini Ultra.^[1]

The report said the smaller models also performed above their weight class. Reka Flash (21B) outperformed GPT-3.5 Turbo and larger models including Grok-1, Mistral Medium, and Gemini Pro 1.0, while Reka Edge (7B) outperformed Gemma 7B and Mistral 7B.^[1]

Reka said Core was trained from scratch on thousands of GPUs over a few months, has a 128K-token context window, and was pretrained on text from 32 languages.^[16] Reka's partners at launch included Snowflake, Oracle, and AI Singapore.^[16]

Oracle partnership (April 2024)

On April 18, 2024, Oracle announced that Reka had selected Oracle Cloud Infrastructure (OCI) to support its work, and that Reka's models, trained on OCI AI Infrastructure powered by Nvidia GPUs, would be made available in the Oracle Cloud Marketplace.^[17] Reka Core and Reka Flash were to be offered through the marketplace via private offers.^[17]

Snowflake acquisition talks (May 2024)

On May 16, 2024, Bloomberg reported that Snowflake was in talks to acquire Reka in a deal that could value it at more than $1 billion, according to people familiar with the matter.^[32] SiliconANGLE noted that the price would be more than three times Reka's reported valuation from its 2023 round, and that Reka had not been generating revenue as of June 2023.^[32] The talks ended without a deal later in May 2024; according to Bloomberg, both companies decided to move on independently but kept their ties.^[5][31]

Partnerships and products (June to December 2024)

On June 4, 2024, Reka announced a multi-year partnership with Shutterstock under which Reka would license Shutterstock's library of images and videos for training, and Shutterstock would use Reka's models to enrich the metadata of its content library.^[38] In August 2024, Reka signed memoranda of understanding with Alibaba Cloud, to offer Reka Core, Flash, and Edge to Alibaba Cloud customers through channels including Model Studio, and with Telkom Indonesia, to develop multimodal models tailored to Bahasa Indonesia and other local languages.^[39][40] In December 2024, Reka introduced Space, a canvas-style interface for its multimodal models powered by Reka Flash, as a free private beta.^[41]

Vibe-Eval benchmark (May 2024)

In May 2024, Reka released Vibe-Eval, an open evaluation benchmark for multimodal chat models.^[7] The benchmark consists of 269 visual understanding prompts, including 100 of hard difficulty, with gold-standard responses written by experts.^[7] More than 50% of the questions in the hard set were answered incorrectly by all frontier models at the time of release.^[7] The evaluation code and data were released publicly, and the work was published as an arXiv preprint titled "Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models."^[7]

Yi Tay's departure (November 2024)

On November 25, 2024, co-founder and chief scientist Yi Tay announced that he was leaving Reka after about 1.5 years to return to Google as a senior staff research scientist, reporting to his former manager Quoc Le.^[6] He wrote, "I still identify more as a scientist / researcher than anything else and hence my decision to return back to my roots."^[6] Looking back on Reka's early work, Tay wrote that Reka Core had debuted on the LMSYS leaderboard in May 2024 with Reka as a top-5 organization, and that "we achieved our goal of outperforming the early versions of gpt-4. IMO that was darn impressive given how much less resources (money, compute and manpower) we had compared to other labs." He added that Reka had raised "a lifetime funding of 100M+ in batches" and that the model was built "with a technical team of less than 15 people."^[6] As of October 2026, his homepage describes him as a principal research scientist on the Gemini core team at Google DeepMind.^[61]

Reka Nexus launch (March 2025)

On March 10, 2025, Reka launched Reka Nexus, an enterprise platform for creating and managing AI workers that automate tasks such as deep topic research, invoice processing, and sales lead generation.^[18] Nexus is built on Reka Flash, which Reka said can be deployed on-premise and on-device with quantization support, and Nexus workers produce human-readable execution traces for auditing.^[18] Yogatama said Reka used Nexus internally for its own sales, recruitment, and operations pipelines.^[18]

Reka Flash 3 open weights release (March 2025)

On March 10, 2025, Reka open-sourced a research preview of Reka Flash 3, a 21-billion-parameter general-purpose reasoning model trained from scratch.^[10] The model was released under the Apache 2.0 license.^[24] Flash 3 was instruction-tuned on curated data and then trained with REINFORCE Leave One-Out (RLOO) using model-based and rule-based rewards.^[10] It marks its thinking with <reasoning> tags, which also allows "budget forcing": cutting off the reasoning after a set number of steps.^[10] The weights were uploaded to Hugging Face under the RekaAI organization in a Llama-compatible format.^[24] Reka described Flash 3 as performing competitively with proprietary models such as OpenAI o1-mini and as "currently the best model in its size category."^[10]

Reka Flash 3.1, released on July 10, 2025, improved by 10 points on LiveCodeBench v5 over Flash 3; Reka said it was competitive on coding with Qwen3-32B and o3-mini and was particularly strong as a base model for fine-tuning on agentic tasks.^[11]

Reka Vision, Reka Research, and Reka Quant (July 2025)

On July 8, 2025, Reka launched Reka Vision, a platform for visual understanding and search built from three modules (Watch, Search, and Chat) that a model planner can combine to automate workflows such as finding moments in long videos, cutting long videos into short social-media clips, and triggering alerts in physical security.^[19] The same day, Reka announced that the security company Turing had launched Guardian AI, an agentic video surveillance product built on Turing's platform and Reka Vision.^[22]

On July 10, 2025, Reka launched Reka Research, an agentic system that browses the web and private documents to answer complex questions; Reka said it was best suited to questions requiring up to 20 sources and typically took one to two minutes to answer.^[42] The same day it released Reka Quant, an open-source quantization library, alongside a 3.5-bit quantized version of Reka Flash 3.1.^[11][20] According to Reka, quantizing Flash 3.1 to the Q3_K_S format in llama.cpp with Reka Quant caused an average performance drop of 1.6 points, compared with 6.8 points for standard Q3_K_S quantization.^[11]

$110 million funding and unicorn status (July 2025)

On July 22, 2025, Reka announced a $110 million investment from new and existing investors including Nvidia and Snowflake.^[8] According to Bloomberg, as reported by SiliconANGLE, the round more than tripled Reka's valuation to about $1 billion, up from about $300 million after its 2023 round.^[3][31] Vivek Raghunathan, Snowflake's vice president of AI engineering, told Bloomberg that Reka would provide AI models and other services to Snowflake, calling Reka "one of the rare independents" that had proven it could compete.^[31] Reka said the money would accelerate technical development and scale its multimodal platforms; SiliconANGLE reported that the company had grown from 20 to 50 employees over the previous year.^[8][31]

Government and defense work (August to October 2025)

In August 2025, Reka and Carahsoft announced a partnership under which Carahsoft would act as Reka's master government aggregator, offering Reka Vision for public safety, Reka Research for government operations, and Reka Speech translation to public-sector agencies.^[44] Reka Speech, announced in September 2025, is an 850-million-parameter model for multilingual speech-to-text and speech-to-speech translation across ten languages.^[43] In October 2025, Reka said it had joined the Oracle Defense Ecosystem, a network of companies working on defense technology for the United States and allied partners.^[45]

Reka Edge for physical AI (March 2026)

On March 11, 2026, Reka released a new version of Reka Edge, a 7-billion-parameter vision language model aimed at physical AI uses such as robots, drones, cars, and cameras.^[21] The model pairs a 657-million-parameter ConvNeXt V2 vision encoder with a 6.4-billion-parameter transformer trained from scratch, and it produces 64 tokens per image tile.^[21] Reka reported that for a 1024 x 1024 image it needs about 3x fewer input tokens than other models of comparable size, and that it processed 5.46 images per second in its tests, more than twice as fast as Cosmos Reason2 8B and Qwen3.5 9B.^[21] With 4-bit quantization, Reka said, memory use drops from 13 GB to about 5 GB while retaining over 98% of the original performance.^[21] The weights are published on Hugging Face as RekaAI/reka-edge-2603 under a custom Reka license that permits commercial use for organizations with less than $1 million in annual revenue.^[26] In April 2026, Reka Edge also became available through OpenRouter.^[47]

Smart City Asia and security focus (May 2026)

In May 2026, Reka showcased its multimodal physical security technology at Smart City Asia 2026 in Ho Chi Minh City, according to TipRanks, which summarized a Reka LinkedIn post.^[28] The post presented Reka Vision as a system for interpreting sensor data and claimed it could speed investigations by up to 20x and cut false alarms by 95%; these are company claims.^[28]

Moonvalley merger (May to June 2026)

On May 10, 2026, the Korean outlet Digital Today, citing The Information and two people familiar with the matter, reported that Reka had acquired Moonvalley, a video-generation startup founded in 2023, in an all-share deal.^[49] The report described Moonvalley as Toronto-based, said it offered video models trained on licensed data to avoid copyright problems, and said it had raised $154 million from investors including General Catalyst, Khosla Ventures, and Y Combinator.^[49] The Information also said Reka had shifted from developing frontier models to enterprise image and video applications and planned to use the deal to expand into world models and robotics, and it presented the deal as an example of how hard it had become for independent model startups to survive.^[49]

Reka announced the combination on its blog on June 9, 2026 and in a press release on June 11, 2026, describing it as having "joined forces with Moonvalley" in a merger that added a team of AI researchers and engineers.^[29][48] The deal brought Mateusz Malinowski (previously a staff research scientist at Google DeepMind) and Mikołaj Bińkowski (previously a senior research scientist at Google DeepMind), both described as key contributors to models that became Google's Veo, into Reka's research leadership, along with former DeepMind, Meta, Amazon, Microsoft, Google, Wayve, and Runway researchers and engineers.^[29] Reka's announcement did not disclose financial terms.^[29]

The combined team's stated focus is a World Language Action Model (WLAM), described as an omni model trained on egocentric and other physical-world data so that it can perceive and act in the real world by running realistic simulations for planning.^[29] Announcing the merger, Yogatama said, "Physical AI is fundamentally a research problem. This team brings world-class expertise in video generation and multimodal models."^[29] Malinowski said, "We're building models that don't just generate video. They understand how the physical world works."^[29]

World-model research releases (June to October 2026)

After the merger, Reka published a series of research releases on world models and physical AI:

In a September 2026 post, Reka argued that large language models were "never the endgame" and that it was building "omni world models" in which a single architecture handles both discrete signals, such as text, and continuous signals, such as images, video, and audio.^[55]

Rho-1 research preview (October 2026)

On October 5, 2026, Reka released a research preview of Rho-1, which it called a 19-billion-parameter "omni-reasoning model" trained from scratch that understands and generates text, images, and video, reasons over them, and takes actions within a single network.^[30] Reka describes two streams, one for understanding and one for generation, that share attention and a KV cache in every block; the model is trained with next-token prediction for discrete sequences and flow matching for continuous image, video, and action outputs. The company reports training the checkpoint from scratch on 320 NVIDIA H100 GPUs for three months.^[30]

Demonstrations include conversational image and video editing, real-time steerable video rollouts, and action output in a LIBERO simulation task. These are company demonstrations, not evidence of general real-world robot reliability. Reka identifies structural drift during long rollouts, unreliable grounding across video, brittle targeted edits, and a native video resolution limit of 672 x 384. It describes Rho-1 as "a functional proof-of-concept and an architectural direction, not a finished product."^[30]

Who founded Reka AI?

TechCrunch's 2023 launch coverage named four co-founders (Dani Yogatama, Cyprien de Masson d'Autume, Qi Liu, and Yi Tay), and Mikel Artetxe also describes himself as a co-founder.^[2][13] The founders had previously worked at Google DeepMind, Google Brain, Meta FAIR, and Baidu. Yogatama leads the company as CEO, and Yi Tay served as chief scientist until November 2024.^[2][6] Reka's 2023 Form D filings list Yogatama, de Masson d'Autume, and Robert Toews as related persons, with Yogatama as an executive officer; TechCrunch identified Rob Toews as a partner at Radical Ventures.^[2][34]

What models does Reka AI make?

Reka has developed models across several size tiers, trained from scratch, including general-purpose multimodal models, reasoning-tuned models, quantized variants, a speech model, a vision language model for physical AI, and, in 2026, a unified understanding-and-generation model.

Reka model timeline

Reka Core

Reka Core is the largest model described in the 2024 technical report, which does not disclose its parameter count.^[1] Reka said Core has a 128K context window, was pretrained on text from 32 languages, and can understand images, video, and audio.^[16]

At release, Core was benchmarked against GPT-4V, Claude 3 Opus, and Gemini Ultra. The version reported in the technical report (Core v0.5) scored 83.2 on MMLU, 92.2 on GSM8K, and 76.8 on HumanEval; on multimodal tasks it scored 56.3 on MMMU, 78.1 on VQAv2, and 59.3 on the Perception-Test video benchmark.^[1]

Reka said Core was available through its API, on-premises, or on-device, and partners announced distribution through Snowflake Cortex and the Oracle Cloud Marketplace.^[15][16][17]

Reka Flash

Reka Flash is a 21-billion-parameter dense model trained on about 5 trillion text tokens, with an 8K standard context window and a 128K long-context version.^[1] Reka called it a "turbo-class" model, offering strong performance at a lower computational cost than Core.^[36]

In the technical report, Flash (v1.5) scored 75.9 on MMLU, 85.8 on GSM8K, 72.0 on HumanEval, and 53.3 on MMMU, ahead of Gemini Pro 1.0 and Llama 2 70B on most of the reported measures.^[1]

Reka Edge

The original Reka Edge is a 7-billion-parameter dense model trained on about 4.5 trillion text tokens, with an 8K standard context window and a 64K long-context version.^[1] It was designed for resource-constrained settings such as on-device and local inference.^[36]

In the technical report, Edge outperformed Mistral 7B and Llama 2 7B on all eight reported benchmarks and Gemma 7B on all except MATH. It scored 65.7 on MMLU, 66.2 on GSM8K, and 54.3 on HumanEval.^[1]

The March 2026 Reka Edge is a new vision language model built for physical AI, with a ConvNeXt V2 vision encoder and a transformer backbone trained from scratch.^[21] Reka said it can be deployed on platforms including NVIDIA Jetson, Apple Silicon Macs, Linux and Windows PCs, and phones and wearables using Qualcomm Snapdragon chips.^[21] In September 2026, Reka released EdgeQ, a version optimized for the Snapdragon 8 Elite; on a Samsung S25, Reka reported EdgeQ scores of 88.65 on RefCOCO-A and 66.93 on MLVU, compared with 73.93 and 32.90 for Google's Gemma 4 E4B run at full precision on a server.^[56]

Reka Flash 3 and Flash 3.1

Released in March 2025 under the Apache 2.0 license, Reka Flash 3 is a 21-billion-parameter general-purpose reasoning model trained from scratch.^[10][24] It has a 32K context window and is released in a Llama-compatible format.^[10][24] Reka said the full-precision model takes 39 GB (fp16) and can be compressed to about 11 GB with 4-bit quantization while maintaining performance, and described it as primarily an English-language model.^[10]

Flash 3.1 followed in July 2025 after what Reka called major upgrades to its reinforcement learning stack, including a new RL algorithm.^[11] A multimodal version of Flash 3.1 serves as the base model for Reka Research and Reka Vision.^[11]

Benchmark performance

The following tables summarize results reported in the April 2024 technical report.^[1] The report labels the evaluated models Reka Core v0.5 and Reka Flash v1.5.

Language benchmarks

Multimodal benchmarks

Medical benchmarks

Human evaluation (multimodal chat ELO)

Yi Tay later wrote that these models were built by "a technical team of less than 15 people," far smaller than the teams at the largest labs.^[6]

Products and services

Reka Chat and Space

At the time of the 2024 technical report, Reka's models were available in a chat playground at chat.reka.ai and through a developer API platform.^[1] In December 2024, Reka introduced Space, a board-style interface for working with its multimodal models, powered by Reka Flash.^[41]

Reka Research

Reka Research is an agentic AI product that browses the web and connected private files (through a Google Drive integration) to answer complex questions, reasoning in natural language before each step.^[42] Reka said it is best suited to factual questions requiring up to 20 sources and typically answers in one to two minutes, with transparent reasoning traces and citations.^[42] A customized version of Reka Flash 3.1 acts as its orchestrator.^[11]

The Reka Research API is compatible with the OpenAI SDK and supports structured JSON output, a cap on reasoning steps, and restricting searches to specific domains.^[42] At launch, Reka priced it at $25 per 1,000 queries, regardless of tokens or steps used.^[42]

Reka Nexus

Reka Nexus is the company's AI workforce platform, launched in March 2025.^[18] It lets organizations create AI workers specialized for tasks such as deep research, invoice processing, or sales lead generation.^[18] Nexus is built on Reka Flash, which can be deployed on-premise or on-device with quantization, and Nexus workers produce human-readable execution traces that support auditing.^[18]

Reka Vision

Reka Vision, launched in July 2025, is a platform for video and image understanding and search.^[19] It is built around three modules: Watch (analyzing uploaded video or images once and storing the result in a memory module), Search (natural-language search for moments, including activities that unfold over time), and Chat (questions and summary reports about the user's media).^[19] A model planner can orchestrate these modules for workflows such as finding specific moments in large video libraries, turning long videos into short clips, and real-time incident alerts.^[19]

Reka said in July 2025 that Reka Vision was used by companies such as Shutterstock and Turing Video.^[8] Turing's Guardian AI product, built on Reka Vision, lets security customers search footage with natural-language queries, set smart alerts, and generate incident reports.^[22] In May 2025, Reka announced an integration of its Reka Vision Agent with NVIDIA AI Blueprint for Video Search and Summarization (VSS).^[23] In November 2025, Reka reported strong results for Reka Vision on SmartHome-Bench, a video anomaly detection benchmark for smart-home footage developed by Wyze.^[46]

Claru

Claru is Reka's data business, which supplies training data for robotics, embodied AI, and world models.^[59] Claru's website, which names Reka AI Inc. as the legal entity, says it works through a network of more than 10,000 data collectors in more than 100 cities.^[59] Reka describes Claru as its "foundational data engine," sourcing egocentric video through paid collectors.^[53]

Enterprise deployment

Reka has offered API access as well as on-premise, VPC, and on-device deployment of its models, and enterprises can customize the models on their own data.^[2][9][16] Its September 2026 terms of service also cover serverless inference and model customization on the Reka Cloud services, billed through prepaid usage credits.^[65]

Partnerships

Did Snowflake acquire Reka AI?

No. Snowflake did not acquire Reka. Snowflake Ventures took part in Reka's 2023 funding round, Snowflake was one of Reka's first customers, and Snowflake backed the $110 million round in 2025.^[2][8] In March 2024, Snowflake added Reka's models to its Cortex AI service.^[4][15] In May 2024, Bloomberg reported that Snowflake was in talks to buy Reka for more than $1 billion, but the talks ended without a deal; according to Bloomberg, both companies decided to move on independently while keeping their ties.^[5][31][32]

Nvidia

Nvidia is an investor in Reka; it was named as a backer of the $110 million round in July 2025.^[8] Reka said in 2025 that it had long used Nvidia's Megatron-Core to train its multimodal models from scratch, and its Reka Vision Agent integrates with NVIDIA AI Blueprint for Video Search and Summarization.^[23] Reka said it worked closely with Nvidia on its real-time video-generation system in 2026, and its Rho-1 checkpoint was trained on Nvidia H100 GPUs.^[30][54]

Oracle

In April 2024, Oracle announced that Reka had selected Oracle Cloud Infrastructure and would offer Reka Core and Reka Flash through the Oracle Cloud Marketplace.^[17] In October 2025, Reka joined the Oracle Defense Ecosystem.^[45]

Shutterstock and Turing Video

In June 2024, Reka signed a multi-year deal to license Shutterstock's visual content for model training, with Shutterstock also becoming a customer that would use Reka models to enrich its library metadata.^[38] In July 2025, Turing launched Guardian AI, an agentic surveillance product built on top of Reka Vision.^[22]

Cloud and telecom partners

In August 2024, Reka signed memoranda of understanding with Alibaba Cloud, covering distribution of Reka models through Alibaba Cloud's Model Studio, and with Telkom Indonesia, covering models for Bahasa Indonesia and local Indonesian languages.^[39][40]

Technical approach

Multimodal from scratch

Reka's 2024 models were trained to process text, images, video, and audio, rather than being text-only models with multimodal features added later.^[1] Its 2026 work goes further: Rho-1 generates as well as understands images and video, and outputs robot actions, within one network that shares a single context window across modalities.^[30]

Training efficiency

Reka has repeatedly emphasized training efficiency. In its 2025 funding announcement, it said its focus on efficient training and serving infrastructure had let it develop models "at a fraction of the cost."^[8] Reka Core was trained on thousands of GPUs over a few months, and the Rho-1 research checkpoint on 320 H100 GPUs for three months, which Reka called "a tiny fraction of the compute behind modern frontier language and video models."^[16][30]

Reinforcement learning

For Reka Flash 3, Reka used REINFORCE Leave One-Out (RLOO) with both model-based and rule-based rewards in the final training stage, focusing on general improvements rather than a model specialized for math or code.^[10] For Flash 3.1, it reported a new RL algorithm and major scalability improvements to its RL stack, and it used the same infrastructure to train the orchestrator of Reka Research end to end.^[11]

Quantization research (Reka Quant)

Reka Quant, released as open source in July 2025, is the company's contribution to model compression.^[20] It combines LDLQ error reduction, which uses precomputed activation statistics to reduce the quantization error of each linear layer, with full-network self-distillation from the original BF16 model while tensors are quantized one at a time; the activation statistics are recomputed online before each quantization step.^[20] The library supports NF4 and GGML (llama.cpp) quantization primitives and exports to GGUF formats.^[27]

Open source contributions

Reka has contributed to the open-source AI ecosystem through several releases. Reka Flash 3 and Flash 3.1 were released under the Apache 2.0 license.^[10][24] The company also released the Vibe-Eval benchmark code and data, the Reka Quant library, the RekaCS2-10k dataset with its open-source renderer, and the RekaDaily-10k dataset, and it published the Reka Inverse Dynamics Model weights under Apache 2.0.^{[7][27][51][53][58]} The 2026 Reka Edge model has open weights under a custom license with a revenue limit for commercial use.^[26]

How is Reka AI funded?

Reka's reported funding consists of a 2023 round of about $58 million to $60 million (reported figures vary) and a $110 million round in July 2025.^[2][8][33] Tay wrote in November 2024 that Reka had raised "a lifetime funding of 100M+ in batches."^[6]

In May 2026, The Information reported that Reka acquired Moonvalley in an all-share deal; no value for that deal was disclosed.^[29][49]

Offices and team

Reka's announcements have given its headquarters as Sunnyvale, California (2025) and as San Francisco (2026), and its legal mailing address is 530 Lawrence Expressway in Sunnyvale.^[8][29][65] At its 2023 launch, Reka described itself as a "globally distributed, highly productive, remote-first team" headquartered in the San Francisco Bay Area.^[33] SiliconANGLE reported in July 2025 that the company had about 50 employees, and its careers page says its team represents 17 nationalities.^[31][64] The Moonvalley merger in 2026 added a further team of researchers and engineers.^[29]

ELI5: Reka AI in simple terms

Reka AI is a small company that builds AI models that can look at pictures and videos, listen to audio, and read text, then answer questions about them. A handful of researchers who used to work at big labs like Google DeepMind and Meta started it in 2022, and they built models that compared well with much bigger labs' models. A big data company called Snowflake talked about buying Reka for about a billion dollars in 2024, but the deal did not happen; in 2025 Nvidia and Snowflake invested in Reka instead, at a value of more than a billion dollars. In 2026 Reka joined up with a video-making AI startup called Moonvalley and started building models that can imagine how the physical world moves, so that robots can plan what to do.

See also

• List of neolabs
• Neolab
• Moonvalley
• Reka Core
• Reka Flash
• Reka Edge
• Large Language Models
• Multimodal AI
• World model
• Physical AI
• Snowflake AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19Reka Team. "Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models." arXiv preprint arXiv:2404.12387 (April 2024). arxiv.org/...2404.12387
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Wiggers, Kyle. "Reka emerges from stealth to build custom AI models for the enterprise." TechCrunch (June 27, 2023). techcrunch.com/...tom-ai-models-for-the-enterprise
3. ^1 ^2 ^3Bloomberg. "Snowflake, Nvidia Back New Unicorn Reka AI in $110 Million Deal." Bloomberg (July 22, 2025). bloomberg.com/...icorn-reka-ai-in-110-million-deal
4. ^1 ^2Sharma, Shubham. "Snowflake teams up with Reka to add multimodal LLMs to data cloud." VentureBeat (March 21, 2024). venturebeat.com/...d-multimodal-llms-to-data-cloud
5. ^1 ^2 ^3Bloomberg. "Snowflake Talks to Acquire Reka AI Said to Fizzle With No Deal." Bloomberg (May 22, 2024). bloomberg.com/...ka-ai-said-to-fizzle-with-no-deal
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7Tay, Yi. "Returning to Google DeepMind." Personal Blog (November 25, 2024). yitay.net/...returning-to-google-deepmind
7. ^1 ^2 ^3 ^4 ^5Reka Team. "Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models." arXiv preprint arXiv:2405.02287 (May 2024). arxiv.org/...2405.02287
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Reka AI. "Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms." Reka Blog (July 22, 2025). reka.ai/...adoption-of-its-multimodal-ai-platforms
9. ^1 ^2 ^3 ^4 ^5Sharma, Shubham. "Reka launches Yasa-1, a multimodal AI assistant to take on ChatGPT." VentureBeat (October 4, 2023). venturebeat.com/...ai-assistant-to-take-on-chatgpt
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Reka AI. "Reasoning with Reka Flash 3." Reka Blog (March 10, 2025). reka.ai/...introducing-reka-flash
11. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Reka AI. "Reka Flash 3.1 and Reka Quant." Reka Blog (July 10, 2025). reka.ai/...reka-flash-3-1-and-reka-quant
12. ^Dani Yogatama. Personal Homepage. dyogatama.github.io
13. ^1 ^2 ^3Mikel Artetxe. Personal Homepage. mikelartetxe.com
14. ^Li, Yujia, et al. (including Cyprien de Masson d'Autume). "Competition-Level Code Generation with AlphaCode." arXiv preprint arXiv:2203.07814 (February 2022). arxiv.org/...2203.07814
15. ^1 ^2 ^3Snowflake. "Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex." Snowflake Blog (March 21, 2024). snowflake.com/...multimodal-llm-snowflake-reka
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Reka AI. "Reka Core: Our Frontier Class Multimodal Language Model." Reka Blog (April 15, 2024). reka.ai/...rontier-class-multimodal-language-model
17. ^1 ^2 ^3 ^4 ^5Oracle Corporation. "Oracle and Reka Collaborate to Advance AI Innovation." PR Newswire (April 18, 2024). prnewswire.com/...-advance-ai-innovation-302120884
18. ^1 ^2 ^3 ^4 ^5 ^6 ^7Reka AI. "Reka launches Nexus, an AI workforce powered by its state-of-the-art multimodal reasoning model." PR Newswire (March 10, 2025). prnewswire.com/...imodal-reasoning-model-302396904
19. ^1 ^2 ^3 ^4 ^5Reka AI. "Reka Vision: Intelligence Made Visible." Reka Blog (July 8, 2025). reka.ai/...reka-vision-intelligence-made-visible
20. ^1 ^2 ^3Reka AI. "Reka Quantization Technology." Reka Blog (July 10, 2025). reka.ai/...reka-quantization-technology
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Reka AI. "Reka Edge: Frontier-Level Edge Intelligence for Physical AI." Reka Blog (March 11, 2026). reka.ai/...level-edge-intelligence-for-physical-ai
22. ^1 ^2 ^3Reka AI. "Reka and Turing Partner to Pioneer Agentic Video Surveillance Platform." Reka Blog (July 8, 2025). reka.ai/...eer-agentic-video-surveillance-platform
23. ^1 ^2Reka AI. "Using NVIDIA AI Blueprint for Video Search and Summarization with Reka Vision Agent." Reka Blog (May 21, 2025). reka.ai/...nd-summarization-with-reka-vision-agent
24. ^1 ^2 ^3 ^4 ^5 ^6Hugging Face. "RekaAI/reka-flash-3." huggingface.co/...reka-flash-3
25. ^Hugging Face. "RekaAI/reka-flash-3.1-rekaquant-q3_k_s." huggingface.co/...reka-flash-3.1-rekaquant-q3_k_s
26. ^1 ^2 ^3Hugging Face. "RekaAI/reka-edge-2603." huggingface.co/...reka-edge-2603
27. ^1 ^2GitHub. "reka-ai/rekaquant." github.com/...rekaquant
28. ^1 ^2TipRanks. "Reka AI Showcases Multimodal Security Platform at Smart City Asia 2026." (May 6, 2026). tipranks.com/...y-platform-at-smart-city-asia-2026
29. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Reka AI. "Reka and Moonvalley Join Forces to Advance Models and Infrastructure for Physical AI." PR Newswire (June 11, 2026). prnewswire.com/...ucture-for-physical-ai-302797869
30. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Reka AI. "Rho-1: Collapsing the multimodal stack." Reka Labs (October 5, 2026). reka.ai/...rho-1-collapsing-the-multimodal-stack
31. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Dotson, Kyt. "Multimodal AI startup Reka AI raises $110M at $1B valuation." SiliconANGLE (July 22, 2025). siliconangle.com/...ka-ai-raises-110m-1b-valuation
32. ^1 ^2 ^3 ^4Deutscher, Maria. "Report: Snowflake in talks to acquire AI model developer Reka AI for $1B+." SiliconANGLE (May 17, 2024). siliconangle.com/...quire-llm-developer-reka-ai-1b
33. ^1 ^2 ^3 ^4 ^5Reka AI. "Announcing our $60M funding to Build Generative Models and Advance AI Research." Reka Blog (June 27, 2023). reka.ai/...nerative-models-and-advance-ai-research
34. ^1 ^2U.S. Securities and Exchange Commission. "Reka AI, Inc. Form D (accession 0001950203-23-000002)." SEC EDGAR (filed July 25, 2023). sec.gov/...primary_doc.xml
35. ^1 ^2 ^3U.S. Securities and Exchange Commission. "Reka AI, Inc. Form D (accession 0001950203-23-000001)." SEC EDGAR (filed July 25, 2023). sec.gov/...primary_doc.xml
36. ^1 ^2 ^3 ^4 ^5 ^6Reka AI. "Reka Flash: Efficient and Capable Multimodal Language Models." Reka Blog (February 12, 2024). reka.ai/...-and-capable-multimodal-language-models
37. ^1 ^2The Decoder. "Yasa-1 is a multimodal AI assistant that aims to compete with ChatGPT." (October 6, 2023). the-decoder.com/...at-aims-to-compete-with-chatgpt
38. ^1 ^2Reka AI. "Reka Announces Partnership with Shutterstock." Reka Blog (June 4, 2024). reka.ai/...announces-partnership-with-shutterstock
39. ^1 ^2Reka AI. "Reka and Alibaba Cloud Announce Partnership to Expand Access to Advanced Multimodal AI." Reka Blog (August 23, 2024). reka.ai/...nnounces-partnership-with-alibaba-cloud
40. ^1 ^2Reka AI. "Reka and Telkom Forge a Strategic Collaboration to Advance AI in Indonesia." Reka Blog (August 28, 2024). reka.ai/...reka-collabs-with-telkom
41. ^1 ^2Reka AI. "Introducing Space." Reka Blog (December 9, 2024). reka.ai/...introducing-space
42. ^1 ^2 ^3 ^4 ^5 ^6Reka AI. "Reka Research: Knowledge Made Accessible." Reka Blog (July 10, 2025). reka.ai/...reka-research-knowledge-made-accessible
43. ^1 ^2Reka AI. "Reka Speech: High Throughput Speech Transcription and Translation Model with Timestamps." Reka Blog (September 16, 2025). reka.ai/...n-and-translation-model-with-timestamps
44. ^Reka AI. "Reka and Carahsoft Partner to Deliver Custom AI Models for Government Agencies." Reka Blog (August 27, 2025). reka.ai/...ustom-ai-models-for-government-agencies
45. ^1 ^2Reka AI. "Reka Joins Oracle's Defense Ecosystem to Advance Multimodal AI for Mission Readiness." Reka Blog (October 14, 2025). reka.ai/...nce-multimodal-ai-for-mission-readiness
46. ^Reka AI. "Reka Vision Sets New Standard for Smart Home & Security AI." Reka Blog (November 20, 2025). reka.ai/...new-standard-for-smart-home-security-ai
47. ^Reka AI. "Switch Models, Zero Code Changes: Reka Edge Now Available on OpenRouter." Reka Blog (April 14, 2026). reka.ai/...s-reka-edge-now-available-on-openrouter
48. ^Reka AI. "Reka and Moonvalley Join Forces to Advance Models and Infrastructure for Physical AI." Reka Blog (June 9, 2026). reka.ai/...dels-and-infrastructure-for-physical-ai
49. ^1 ^2 ^3 ^4 ^5Hwang, Chi-gyu. "Reka AI seeks to expand world model and robotics business with Moonvalley acquisition." Digital Today (May 10, 2026). digitaltoday.co.kr/...-with-moonvalley-acquisition
50. ^Reka AI. "PhysicalRealismBench-U: Attributable Physical Realism Evaluation for Video World Models." Reka Blog (June 9, 2026). reka.ai/...alism-evaluation-for-video-world-models
51. ^1 ^2Reka AI. "RekaCS2-10k: A Large-Scale Egocentric Counter-Strike 2 Dataset." Reka Blog (June 24, 2026). reka.ai/...ale-egocentric-counter-strike-2-dataset
52. ^Reka AI. "WorldModelGym: a decision-based fidelity benchmark for world models." Reka Blog (July 2, 2026). reka.ai/...worldmodelgym
53. ^1 ^2 ^3Reka AI. "RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data." Reka Blog (August 6, 2026; updated September 16, 2026). reka.ai/...-egocentric-household-manipulation-data
54. ^1 ^2Reka AI. "Real-Time Video Generation." Reka Blog (August 14, 2026). reka.ai/...real-time-video-generation
55. ^Reka AI. "Evolution of LLMs: Omni-World Models." Reka Blog (September 17, 2026). reka.ai/...evolution-of-llms-omni-world-models
56. ^1 ^2Reka AI. "Reka EdgeQ, On-Device." Reka Blog (September 22, 2026). reka.ai/...reka-edgeq
57. ^1 ^2Reka AI. "Reka Inverse Dynamics Model for Interactive World Models." Reka Blog (October 2, 2026). reka.ai/...mics-model-for-interactive-world-models
58. ^1 ^2 ^3Hugging Face. "RekaAI/Reka-Inverse-Dynamics-Model." huggingface.co/...Reka-Inverse-Dynamics-Model
59. ^1 ^2Claru. "Claru: Training Data for Physical AI." claru.ai
60. ^Reka AI. Homepage. reka.ai
61. ^1 ^2Yi Tay. Personal Homepage. yitay.net
62. ^Qi Liu. Personal Homepage. leuchine.github.io
63. ^Rae, Jack W., et al. (including Cyprien de Masson d'Autume). "Scaling Language Models: Methods, Analysis & Insights from Training Gopher." arXiv preprint arXiv:2112.11446 (December 2021). arxiv.org/...2112.11446
64. ^Reka AI. "Careers." reka.ai/careers
65. ^1 ^2Reka AI. "Terms of Use." (last updated September 2026). reka.ai/...terms-of-use