# Reka AI

> Source: https://aiwiki.ai/wiki/reka_ai
> Updated: 2026-06-27
> Categories: AI Companies, Large Language Models, Multimodal AI
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

Reka AI (commonly referred to as Reka) is an artificial intelligence research and product company, founded in 2022 by former [Google DeepMind](/wiki/google_deepmind), Google Brain, Meta FAIR, and Baidu researchers, that builds natively [multimodal](/wiki/multimodal) [large language models](/wiki/large_language_model) able to process text, images, video, and audio in a single unified model.[1][2] 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 reached unicorn status in July 2025 when [Nvidia](/wiki/nvidia) and [Snowflake](/wiki/snowflake_ai) backed a $110 million round that valued the company at over $1 billion.[3] A reported 2024 bid by Snowflake to acquire Reka for more than $1 billion did not close.[5]

Reka is headquartered in Sunnyvale, California, and maintains operations in Singapore, with team members distributed across London, Zurich, Seattle, and Hong Kong. The company emerged from stealth in June 2023 with $58 million in funding and has since released a family of models spanning different size classes: Reka Core, Reka Flash, and Reka Edge.[2] In July 2025, Reka raised $110 million in a funding round backed by Nvidia and Snowflake, reaching a valuation of over $1 billion and achieving unicorn status.[3] The company has expanded beyond foundation models into product offerings such as Reka Nexus (an AI workforce platform), Reka Vision (a video and image understanding system), and Reka Research (an agentic web research tool), and in June 2026 it merged with the physical-AI startup Moonvalley to push into world models and robotics.[29]

## What is Reka AI?

Reka AI is an enterprise-focused frontier AI lab that designs multimodal foundation models from scratch, meaning a single model is trained on text, images, video, and audio together rather than bolting separate encoders onto a text-only base.[1] The company differentiates itself on efficiency (building near-frontier models with a team of roughly 20 to 50 researchers), on flexible deployment (API, on-premise, and virtual private cloud via Docker containers), and on open releases such as Reka Flash 3 under the [Apache 2.0 license](/wiki/apache_license).[2][10] Reka sells both raw models and higher-level products: Reka Nexus (an AI workforce platform), Reka Vision (multimodal video and image understanding), and Reka Research (agentic web and document research).

## History

### Founding and early development

Reka was founded in 2022 by five AI researchers who had previously worked at some of the world's leading AI laboratories.[2] The founders, Dani Yogatama, Yi Tay, Cyprien de Masson d'Autume, Qi Liu, and Mikel Artetxe, shared a conviction that it was impractical to expect a single large language model to serve all possible enterprise use cases.[2] While working on projects like [AlphaCode](/wiki/alphacode) and Google's [Bard](/wiki/bard) at DeepMind and Google, the founders observed that general-purpose models often fell short when organizations needed AI tailored to specific requirements, such as generating marketing copy in a particular brand voice or processing domain-specific documents.[2]

The company operated in stealth mode through the first half of 2023, assembling a small research team and beginning work on its first models.

### Emergence from stealth (June 2023)

On June 28, 2023, Reka publicly emerged from stealth, announcing it had raised $58 million in a funding round led by DST Global Partners and Radical Ventures, with participation from Snowflake Ventures and several angel investors including former GitHub CEO Nat Friedman.[2] At the time of this announcement, the company was valued at approximately $300 million.[2]

Reka's pitch to investors centered on its ability to build efficient, customizable AI models for enterprise deployment.[2] Unlike larger competitors that focused primarily on consumer-facing chatbots, Reka emphasized on-premise and virtual private cloud (VPC) deployment options that would allow enterprises to keep sensitive data within their own infrastructure.[2]

### Yasa-1 launch (October 2023)

In October 2023, Reka launched Yasa-1, its first publicly available multimodal AI assistant.[9] Yasa-1 was built on a unified model trained from scratch and could process text, images, short videos, and audio snippets.[9] The assistant supported 20 languages and offered a context window of up to 24,000 tokens by default, with the ability to handle documents as long as 100,000 tokens.[9]

Yasa-1 was made available through APIs and as Docker containers for on-premise or VPC deployment.[9] The system could also be customized on private datasets of any modality, allowing enterprises to fine-tune the model for specific use cases.[9]

### Reka Flash release (November 2023)

Following the Yasa-1 launch, Reka released Reka Flash, a 21 billion parameter multimodal language model that served as the company's mid-tier "turbo-class" offering. Flash was designed to deliver strong performance at a fraction of the computational cost required by larger models. It could process text, images, and video inputs and was positioned as a practical choice for everyday enterprise applications.

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

On April 15, 2024, Reka published a comprehensive technical report detailing its full model lineup: Reka Core, Reka Flash, and Reka Edge.[1] The report, titled "Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models," was released as a preprint on arXiv and demonstrated that a small team could build frontier-competitive models.[1]

Reka Core, the company's most capable model, approached the performance of [GPT-4](/wiki/gpt-4)V on image question-answering benchmarks and ranked as the second most preferred model (behind only GPT-4V) in blind third-party human evaluations of multimodal chat, outperforming [Claude](/wiki/claude) 3 Opus.[1] On video question answering, Core surpassed [Gemini](/wiki/gemini) Ultra on the Perception-Test benchmark.[1]

The smaller models also performed well above their weight class. Reka Flash (21B parameters) outperformed much larger models including Gemini Pro 1.0 and [Llama](/wiki/llama) 2 70B on multiple benchmarks, while Reka Edge (7B parameters) beat other models in its size category, including [Mistral](/wiki/mistral) 7B and [Gemma](/wiki/gemma) 7B.[1]

Reka Core was released with launch API pricing of $10 per million input tokens and $25 per million output tokens, putting it in the same general price tier as competing frontier models at the time.[16] The model was pretrained on textual data covering 32 languages, with fluency demonstrated in English and several Asian and European languages.[1]

### Oracle partnership (April 2024)

On April 18, 2024, only days after the Reka Core release, Reka announced a partnership with Oracle Corporation.[17] Under the agreement, Reka selected Oracle Cloud Infrastructure (OCI) as its preferred cloud provider for training and serving its multimodal models, running on OCI AI Infrastructure with [Nvidia](/wiki/nvidia) GPUs.[17] Reka Core and Reka Flash were also added to the Oracle Cloud Marketplace, giving Oracle's enterprise customers a way to deploy Reka models alongside Oracle databases and applications.[17]

### Snowflake acquisition talks and partnership (May 2024)

In May 2024, Bloomberg reported that [Snowflake](/wiki/snowflake_ai) was in talks to acquire Reka for more than $1 billion.[5] The reporting framed the discussion as part of Snowflake's push to compete with [Databricks](/wiki/databricks) and other data platforms that were rapidly absorbing or building generative AI capabilities.[5] By May 22, 2024, those talks had broken down without a deal, with both companies deciding it made more sense to move forward independently.[5] Bloomberg reported that the parties could not agree on price and terms, and that Reka's existing investors and founders preferred to keep the company independent.[5]

Despite the failed acquisition, Reka and Snowflake deepened their partnership rather than walking away. Snowflake integrated Reka Flash and Reka Core into its Cortex AI service, making them available to over 400 enterprises using Snowflake's data cloud.[4] This integration enabled customers to build generative AI applications that could work with text, images, and video inputs directly within Snowflake's platform.[15] Snowflake Ventures, which had been an investor since the 2023 round, kept its stake and later returned in the 2025 growth round.[3]

### Vibe-Eval benchmark (May 2024)

In May 2024, Reka released Vibe-Eval, an open evaluation benchmark for multimodal language models.[7] The benchmark consists of 269 visual understanding prompts, including 100 prompts rated as hard difficulty, with gold-standard responses authored by human experts.[7] Over 50% of the questions in the hard set were answered incorrectly by all frontier models at the time of release, making it a useful tool for measuring real progress in multimodal AI.[7] The benchmark and its dataset were released on GitHub under an open license, 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 his departure from Reka to return to [Google DeepMind](/wiki/google_deepmind) as a Senior Staff Research Scientist.[6] Tay had spent approximately 1.5 years at Reka and cited his identity as a researcher and scientist as the primary reason for his return, writing, "I still identify more as a scientist / researcher than anything else and hence my decision to return back to my roots."[6] Reflecting on what the small Reka team had achieved, Tay wrote, "We achieved our goal of outperforming the early versions of gpt-4. IMO that was darn impressive given how much less resources we had compared to other labs."[6] At DeepMind, he went on to contribute to work on the Gemini family of models.[6]

### Reka Nexus launch (March 2025)

On March 10, 2025, Reka launched Reka Nexus, an AI platform aimed at enterprise customers that lets organizations create and manage AI workers to automate tasks like deep topic research, invoice processing, and sales lead generation.[18] Nexus is built on top of Reka Flash and presents itself as a workforce layer rather than a raw model API.[18] Each AI worker can be customized for a specific role, deployed on-premise or on-device with quantization support, and produces a human-readable execution trace that customers can use for auditing.[18]

### Reka Flash 3 open weights release (March 2025)

On March 11, 2025, Reka open-sourced 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](/wiki/apache_license), making it freely available for commercial and research use.[10] Flash 3 was trained on synthetic and public datasets using supervised fine-tuning, followed by RLOO (REINFORCE Leave One-Out) with model-based and rule-based rewards.[10] The model uses reasoning tags to mark the boundaries of its thinking process, similar to the approach in OpenAI's o1 series, and the weights were uploaded to Hugging Face under the RekaAI organization.[24] Reka described Flash 3 as "currently the best model in its size category."[10]

Reka Flash 3 performed competitively with proprietary models such as [OpenAI](/wiki/openai) o1-mini, which was notable given its relatively small parameter count.[10] An updated version, Reka Flash 3.1, followed shortly after, improving by 10 points on [LiveCodeBench](/wiki/livecodebench) v5 and showing particular strength on coding tasks and as a base model for agentic fine-tuning.[11]

### Reka Vision launch (July 2025)

On July 8, 2025, Reka introduced Reka Vision, a multimodal video and image understanding platform aimed at content creators, security operators, and media companies.[19] Reka Vision is structured around three modules: Watch, Search, and Chat, which can be combined by a model planner to handle workflows like clipping highlights from long videos, finding specific moments inside large media libraries, and triggering alerts when a security camera sees something unusual.[19]

Early customers included Shutterstock, which used Reka Vision to enrich its image and video catalog with natural-language search, and Turing Video, which built an agentic surveillance product called Guardian AI on top of the platform.[22] The system targets long-form video understanding, where each frame can be searched, summarized, or used as the trigger for an alert without sending data to a third party.[19]

### Reka Quant and Flash Quantized release (July 2025)

On July 10, 2025, Reka released Reka Quant, an open-source model quantization library based on its internal toolchain, alongside a 3.5-bit quantized version of Reka Flash 3.1.[20] Reka Quant uses calibrated error reduction and online self-distillation to keep quality close to the original 16-bit model.[20] According to the company, quantizing Flash 3.1 to the Q3_K_S format in llama.cpp using Reka Quant produced an average performance drop of only 1.6 points across benchmarks, compared to 6.8 points using standard Q3_K_S quantization.[20] The library supports NF4 and GGML quantization primitives, distributed Hessian computation for fast LDLQ, and exporting to common llama.cpp formats including Q3_K and Q4_K.[27]

### $110 million funding and unicorn status (July 2025)

In July 2025, Reka raised $110 million in a funding round backed by Nvidia and Snowflake.[3] This round more than tripled the company's valuation from approximately $300 million to over $1 billion, making Reka a unicorn.[3] The investment reflected confidence in Reka's ability to develop market-leading models at a fraction of the cost incurred by larger competitors.[8] Bloomberg's coverage of the round noted that both Snowflake and Nvidia were doubling down on a company they had previously considered acquiring or had supported through earlier rounds, and that the new capital would be used to expand the team and accelerate the rollout of Nexus, Vision, and the next generation of foundation models.[3]

### Reka Edge for Physical AI (March 2026)

On March 11, 2026, Reka announced a new generation of Reka Edge, a 7 billion parameter vision language model aimed at physical AI use cases such as robotics, edge cameras, and embedded inspection systems.[21] The 2026 Reka Edge variant uses 3x fewer input tokens than competing 8B models on the same image-and-text inputs and achieves up to 65% higher throughput.[21] With 4-bit quantization, memory consumption drops from 13 GB to about 5 GB, a 62% reduction, while retaining over 98% of the original benchmark performance and delivering up to 2.3x faster inference.[21] The model targets image understanding, video analysis, object detection, and agentic tool-use, and is published on Hugging Face as `RekaAI/reka-edge-2603`.[26]

### Smart City Asia and security focus (May 2026)

In May 2026, Reka showcased its multimodal physical security platform at Smart City Asia 2026 in Ho Chi Minh City.[28] The demonstration centered on Reka Vision running on top of Reka Edge for live video analysis at scale, and on the company's pitch that a single multimodal model could replace stacks of narrow vision systems used in traffic, retail, and public safety deployments.[28]

### Moonvalley merger (June 2026)

On June 11, 2026, Reka announced that it had joined forces with Moonvalley, a Toronto-based video-generation startup founded in 2023, to accelerate models and infrastructure for physical AI.[29] Moonvalley had raised about $154 million from investors including General Catalyst, Khosla Ventures, and Y Combinator, and was known for video models trained on licensed data.[29] The deal brought Mateusz Malinowski (previously a Staff Research Scientist at Google DeepMind) and Mikolaj Binkowski (previously a Senior Research Scientist at Google DeepMind), both key contributors to Google's Veo video model, into Reka's research leadership along with a team of former DeepMind, Meta, Amazon, Microsoft, Google, Wayve, and Runway researchers.[29] No financial terms were disclosed.[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 it can perceive and act in the real world through realistic simulation for planning.[29] Announcing the merger, Reka CEO Dani Yogatama said, "Physical AI is fundamentally a research problem. This team brings world-class expertise in video generation and multimodal models."[29] Mateusz Malinowski added, "We're building models that don't just generate video. They understand how the physical world works."[29]

## Who founded Reka AI?

Reka was founded by five researchers with complementary backgrounds in natural language processing, [machine learning](/wiki/machine_learning), and large-scale systems engineering. The team had previously worked at Google DeepMind, Google Brain, Meta FAIR, and Baidu, and the company is led by co-founder and CEO Dani Yogatama, with Yi Tay serving as Chief Scientist until November 2024.[2][6]

| Founder | Role at Reka | Previous Affiliation | Background |
|---|---|---|---|
| Dani Yogatama | CEO and Co-founder | [DeepMind](/wiki/google_deepmind) (2016-2022), Baidu Silicon Valley AI Lab (2015-2016) | PhD in Computer Science from Carnegie Mellon University. Senior Staff Research Scientist at DeepMind. Associate Professor at the University of Southern California.[12] |
| Yi Tay | Co-founder, Chief Scientist (until Nov 2024) | Google Brain, [DeepMind](/wiki/google_deepmind) | Co-lead of [PaLM](/wiki/palm) 2 at Google. Inventor of UL2 and Differentiable Search Indexes. Key contributor to Flan-T5 and other instruction-tuning work. Returned to Google DeepMind in November 2024.[6] |
| Cyprien de Masson d'Autume | CTO and Co-founder | [DeepMind](/wiki/google_deepmind) (2016-2022) | Staff Research Engineer at DeepMind. Worked on Gopher and AlphaCode.[14] |
| Qi Liu | Co-founder | DeepMind, Meta FAIR, Microsoft Research | PhD from the University of Oxford. Assistant Professor at the University of Hong Kong. |
| Mikel Artetxe | Co-founder | Meta FAIR | PhD from the University of the Basque Country. Research focus on multilingual NLP, unsupervised machine translation, and cross-lingual representation learning. Honorary Researcher at the University of the Basque Country.[13] |

## What models does Reka AI make?

Reka has developed models across multiple size tiers, all trained from scratch with native multimodal capabilities. The full lineup includes general-purpose multimodal models, reasoning-tuned models, quantized variants, and specialized vision language models for physical AI.

### Reka model timeline

| Model | Release | Parameters | Modalities | Notes |
|---|---|---|---|---|
| Yasa-1 | October 2023 | Undisclosed | Text, image, short video, audio | First public Reka product. 20 languages. 24K default context, up to 100K.[9] |
| Reka Flash 1.0 | November 2023 | 21B | Text, image, video | Original turbo-class multimodal model. |
| Reka Edge 1.0 | April 2024 | 7B | Text, image, video | Compact frontier model for resource-constrained deployment.[1] |
| Reka Core | April 15, 2024 | ~67B (reported) | Text, image, video, audio | Frontier-class multimodal model. 128K context. Trained on 32 languages.[16] |
| Reka Flash 3 | March 11, 2025 | 21B | Text, image | Open weights under Apache 2.0. Reasoning-tuned with RLOO. 32K context.[10] |
| Reka Flash 3.1 | 2025 | 21B | Text, image | +10 points on LiveCodeBench v5. Stronger coding and agentic baseline.[11] |
| Reka Spark | 2025 | ~2B | Text, image | Ultra-compact tier for lightweight tasks. |
| Reka Flash 3.1 RekaQuant Q3_K_S | July 10, 2025 | 21B (3.5-bit) | Text, image | Quantized version. 1.6-point average drop vs full precision.[25] |
| Reka Edge 2603 | March 11, 2026 | 7B | Image, video, text | New vision language model for physical AI. 4-bit quantized variant uses about 5 GB memory.[21] |

### Reka Core

Reka Core is the company's largest and most capable model. While Reka has not officially disclosed the exact parameter count, reports suggest it has approximately 67 billion parameters. Core features a 128K context window and can process text, images, video, and audio inputs simultaneously.[16]

Core was designed as a frontier-class model and was benchmarked against GPT-4V, Claude 3 Opus, and Gemini Ultra at the time of its release. On language tasks, it achieved an MMLU score of 83.2, a GSM8K score of 92.2, and a [HumanEval](/wiki/humaneval) score of 76.8. On multimodal tasks, it scored 56.3 on [MMMU](/wiki/mmmu), 78.1 on VQAv2, and 59.3 on the Perception-Test video benchmark.[1]

Reka Core launched at $10 per million input tokens and $25 per million output tokens, available through Reka's own API, through Snowflake Cortex, through the Oracle Cloud Marketplace, and via on-premise deployment using Docker containers.[16]

### Reka Flash

Reka Flash is a 21 billion parameter dense model trained on approximately 5 trillion text tokens with an 8K default context window that can extend to 128K.[1] Flash was positioned as a "turbo-class" model, offering fast inference and strong performance at a lower computational cost than Core.

Flash outperformed several larger models on key benchmarks: it scored 75.9 on [MMLU](/wiki/mmlu), 85.8 on [GSM8K](/wiki/gsm8k), 72.0 on HumanEval, and 53.3 on MMMU. These scores surpassed Gemini Pro 1.0 and Llama 2 70B on multiple measures.[1]

### Reka Edge

Reka Edge is a 7 billion parameter dense model trained on 4.5 trillion text tokens. It was designed for resource-constrained environments such as on-device deployment and local inference scenarios. Edge features an 8K default context window that can extend to 64K.[1]

Despite its compact size, Edge outperformed other models in the 7B class, including Mistral 7B, Gemma 7B, and Llama 2 7B. It scored 65.7 on MMLU, 66.2 on GSM8K, and 54.3 on HumanEval.[1]

In March 2026, Reka released a new generation of Reka Edge specifically engineered for physical AI applications.[21] The 2026 release uses 3x fewer input tokens and reaches 65% higher throughput than leading 8B models in the same class.[21] With 4-bit quantization, memory drops from 13 GB to about 5 GB while retaining 98% of original performance, and the model can process about 5.46 images per second on standard inference hardware.[21] Reka pitched this version as the model that lets robots, drones, and security cameras run a real frontier-style vision language model on the device itself rather than streaming everything to the cloud.[21]

### 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] Unlike the earlier Flash model, Flash 3 was specifically optimized for reasoning tasks through reinforcement learning (RLOO).[10] It uses a Llama-compatible architecture and a 32K context window.[24] The reasoning trace is wrapped in dedicated tags so that downstream applications can choose to display, hide, or audit the model's intermediate steps.[10]

Flash 3.1, an updated version released shortly after, incorporated advances in Reka's reinforcement learning stack and improved by 10 points on LiveCodeBench v5.[11] Flash 3.1 showed particular strength on coding benchmarks and as a base model for fine-tuning on agentic tasks.[11]

In July 2025, Reka released a 3.5-bit quantized variant of Flash 3.1 produced with the open-source Reka Quant library.[20] The quantized variant runs in roughly the same memory budget as a 7B model in full precision, while keeping benchmark loss to about 1.6 points on average.[25]

### Reka Spark

Reka Spark is an ultra-compact model with approximately 2 billion parameters, designed for lightweight tasks and edge deployment on smaller devices. Spark serves as the most affordable entry point in Reka's model lineup.

## Benchmark performance

The following table summarizes Reka's model performance on key benchmarks alongside leading competitors, as reported in the April 2024 technical report.[1]

### Language benchmarks

| Benchmark | Reka Edge (7B) | Reka Flash (21B) | Reka Core | GPT-4 (0613) | Claude 3 Opus | Gemini Ultra |
|---|---|---|---|---|---|---|
| [MMLU](/wiki/mmlu) | 65.7 | 75.9 | 83.2 | 86.4 | 86.8 | 83.7 |
| [GSM8K](/wiki/gsm8k) | 66.2 | 85.8 | 92.2 | 92.0 | 95.0 | 94.4 |
| [HumanEval](/wiki/humaneval) | 54.3 | 72.0 | 76.8 | 76.5 | 84.9 | 74.4 |
| GPQA | - | 34.0 | 38.2 | 38.1 | 50.2 | 35.7 |

### Multimodal benchmarks

| Benchmark | Reka Flash (21B) | Reka Core | GPT-4V | Claude 3 Opus | Gemini Ultra |
|---|---|---|---|---|---|
| [MMMU](/wiki/mmmu) | 53.3 | 56.3 | 56.8 | 59.1 | 59.4 |
| VQAv2 | 78.4 | 78.1 | 77.2 | - | 77.8 |
| Perception-Test (Video) | 56.4 | 59.3 | - | - | 54.7 |

### Medical benchmarks

| Benchmark | Reka Edge | Reka Flash | Reka Core | GPT-4 |
|---|---|---|---|---|
| MedMCQA | 52.6 | 71.3 | 80.6 | 72.4 |
| PubMedQA | 71.6 | 69.0 | 74.6 | 80.4 |
| MMLU Medical | 65.7 | 79.5 | 88.3 | 90.3 |

### Human evaluation (multimodal chat ELO)

| Model | ELO Score | Win Rate |
|---|---|---|
| GPT-4V | 1201 | 79.4% |
| Reka Core | 1130 | 72.2% |
| Reka Flash | 1082 | 66.8% |
| Claude 3 Opus | 1073 | 66.2% |

These results were notable because Reka built these models with a team of roughly 20 core researchers, while competitors employed teams that were orders of magnitude larger.[1]

### Reasoning and coding benchmarks (Flash 3 and Flash 3.1)

| Benchmark | Reka Flash 3 | Reka Flash 3.1 | OpenAI o1-mini |
|---|---|---|---|
| [LiveCodeBench](/wiki/livecodebench) v5 | baseline | +10 over Flash 3 | competitive |
| AIME (math) | competitive with o1-mini | improved | reference |
| MMLU-Pro reasoning | competitive | competitive | reference |

The reasoning-tuned Flash 3 family closed much of the gap to closed-source reasoning models in math, code, and tool-use evaluations, despite weighing in at only 21B parameters and being released with open weights under Apache 2.0.[10]

## Products and services

### Reka Chat

Reka Chat is the company's consumer-facing chat interface, available at chat.reka.ai. It allows users to interact with Reka's models for text conversations, document analysis, and multimodal queries involving images and files.

### Reka Research

Reka Research is an agentic AI product that can browse the web and search through private documents to answer complex questions. The tool performs multiple web searches, visits dozens of websites in one to three minutes, and synthesizes information from multiple sources in a multi-hop manner, reasoning in natural language before taking each step. Every answer includes a traceable path of the steps taken.

The Reka Research API is OpenAI-compatible and supports customizable execution, including domain restrictions and structured JSON output. Enterprises can deploy Reka Research on-premise, in a private cloud, or through the API.

### Reka Nexus

Reka Nexus is the company's AI workforce platform, launched in March 2025.[18] It allows enterprise customers to spin up specialized AI workers that handle defined business tasks, such as conducting deep research on a topic, processing invoices, or generating sales leads.[18] Each Nexus worker is built on top of Reka Flash and can be customized with the customer's own data, deployed on-premise or on-device, and quantized for cost-sensitive workloads.[18] The platform produces a human-readable trace of every action a worker takes, which is useful in regulated industries that need an audit log for any AI-driven decision.[18]

### Reka Vision

Reka Vision, launched in July 2025, is a multimodal video and image understanding platform.[19] The product is built around three primitives: Watch (continuous monitoring of video streams or libraries), Search (natural-language retrieval over images and clips), and Chat (multimodal Q&A grounded in the user's media).[19] A model planner orchestrates these primitives to handle higher-level workflows, such as turning a long YouTube video into a set of social-media-ready short clips, or scanning thousands of hours of camera footage for a specific incident.[19]

Early Reka Vision deployments include Shutterstock, which uses the platform to power semantic search across its image and video catalog, and Turing Video, which built Guardian AI, an agentic surveillance product, on top of the platform.[22] Reka Vision is also integrated with Nvidia's AI Blueprint for video search and summarization, allowing customers to combine Reka's reasoning with Nvidia's reference video pipeline.[23]

### Enterprise deployment

Reka offers flexible deployment options for enterprises, including API access, on-premise deployment, and private cloud (VPC) deployment via Docker containers.[2] This flexibility is a key differentiator, as many competing AI providers only offer cloud-based API access. Enterprises can also fine-tune Reka's models on their own proprietary datasets.

### Pricing

| Tier | Use case | Notes |
|---|---|---|
| Reka Core API | Frontier multimodal use | $10 per million input tokens, $25 per million output tokens at launch.[16] |
| Reka Flash API | Mid-tier multimodal use | Lower per-token price than Core. Available via Reka API, Snowflake Cortex, and Oracle Cloud Marketplace. |
| Reka Edge / Spark | Edge and on-device | Open weights for the reasoning Flash 3 family. Edge can be deployed on a single GPU or quantized for CPU. |
| On-premise / VPC | Regulated and large enterprise | Docker container licensing. Pricing negotiated per deployment. |
| Reka Nexus | AI workforce | Per-seat or per-task licensing for AI workers. |
| Reka Vision | Video and image understanding | Usage-based pricing for video minutes processed and image queries. |

## Partnerships

### Did Snowflake acquire Reka AI?

No. Snowflake did not acquire Reka. Snowflake has been one of Reka's most significant partners and investors: Snowflake Ventures participated in Reka's initial $58 million funding round in 2023 and later backed the $110 million round in 2025.[3] In May 2024, Bloomberg reported that Snowflake and Reka were in advanced talks for a full acquisition valued at over $1 billion, but those talks ended on May 22, 2024 without a deal.[5] Rather than walk away, the companies turned the discussion into a deeper commercial partnership. Reka Flash and Reka Core were integrated into Snowflake's Cortex AI service, making Reka's multimodal capabilities available to over 400 enterprises using Snowflake's data cloud, and Snowflake Ventures kept its equity stake.[4]

The integration allows Snowflake customers to build generative AI applications that can process text, images, and video inputs without moving data outside of Snowflake's platform.[15] Reka's models brought the total number of LLMs available in Snowflake Cortex to approximately a dozen.[15]

### Nvidia

Nvidia has been both an investor and a technology partner for Reka. Beyond participating in the $110 million funding round in 2025, Nvidia's GPU infrastructure serves as the primary compute platform for training and serving Reka's models.[3] The partnership reflects the close relationship between frontier AI model developers and Nvidia's hardware ecosystem. Reka Vision is also integrated with Nvidia's AI Blueprint for video search and summarization, which packages a reference architecture for ingesting video, extracting captions, and answering natural-language queries.[23]

### Oracle

In April 2024, Reka announced a partnership with Oracle.[17] Reka selected Oracle Cloud Infrastructure (OCI) as its preferred cloud platform for training and serving multimodal models, running on OCI's Nvidia GPU clusters.[17] Reka Core and Reka Flash were also added to the Oracle Cloud Marketplace, where Oracle's enterprise customers can deploy them alongside Oracle databases, ERP systems, and analytics tooling.[17] The partnership gave Reka a third major distribution channel beyond its own API and the Snowflake Cortex integration.

### Shutterstock and Turing Video

Following the launch of Reka Vision in July 2025, Reka announced anchor customers in two adjacent markets.[19] Shutterstock adopted Reka Vision for natural-language search across its image and video catalog, allowing users to find clips by describing the content rather than guessing tags.[19] Turing Video built Guardian AI, an agentic surveillance product, on top of Reka Vision, with capabilities including searching for specific events across recorded footage, configuring smart alerts, and generating incident summaries written in natural language.[22]

## Technical approach

### Multimodal from scratch

Unlike many competitors that built multimodal capabilities as extensions to existing text-only models, Reka trained its models as natively multimodal from the beginning.[1] This approach means the models can process and reason across text, images, video, and audio within a single unified architecture, rather than relying on separate encoders bolted onto a text model after the fact.

This design choice has direct consequences for what the models can do. Reka Core can take an image and a long audio clip in the same prompt and reason about both together, instead of running speech-to-text first and feeding only the transcript to a language model. The same logic applies to video, where Reka's models work directly on frame-level features rather than receiving a separate caption pipeline as input.

### Training efficiency

One of Reka's defining characteristics is its emphasis on training efficiency. The company has consistently achieved competitive results with significantly fewer resources than larger competitors. The original Reka Core model was developed by a team of roughly 20 researchers, compared to the hundreds or thousands of engineers at organizations like [OpenAI](/wiki/openai), [Google](/wiki/google), and [Anthropic](/wiki/anthropic).[1] Reflecting on this, co-founder Yi Tay wrote that the team "achieved our goal of outperforming the early versions of gpt-4" and called the result "darn impressive given how much less resources we had compared to other labs."[6]

### Reinforcement learning with RLOO

For its 2025 reasoning models, Reka adopted REINFORCE Leave One-Out (RLOO) as its main reinforcement learning algorithm rather than PPO or DPO.[10] RLOO uses multiple sampled responses for each prompt and treats one held-out sample as the baseline for the others, which sidesteps the cost and instability of training a separate value network. Reka combines RLOO with both rule-based rewards (for tasks like math and code where correctness can be checked automatically) and model-based rewards (for tasks like instruction following).[10] The reasoning trace is part of the training signal, so the model learns when a longer chain of thought helps and when it does not.

### Quantization research (Reka Quant)

Reka Quant, released as open source in July 2025, is the company's contribution to the model compression literature.[20] The library combines two ideas: calibrated error reduction, which uses a small calibration set to learn weight-update directions that minimize quantization loss, and online self-distillation, which uses the original model as a teacher during quantization-aware fine-tuning.[20] Together, these techniques let Reka quantize Flash 3.1 to 3.5 bits per weight while losing only about 1.6 average benchmark points, compared to 6.8 points using the standard Q3_K_S baseline in llama.cpp.[20]

### 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, making them freely available for commercial and research use.[10] The company also released the Vibe-Eval benchmark as an open resource for the research community, providing a standardized way to evaluate multimodal language models with expert-annotated gold-standard responses.[7] The Reka Quant library was released on GitHub under an open license in July 2025, and the 4-bit quantized 2026 Reka Edge model was published on Hugging Face for local deployment.[27]

## How is Reka AI funded?

Reka has raised approximately $168 million in disclosed funding across two rounds: a $58 million Series A in June 2023 at roughly a $300 million valuation, and a $110 million growth round in July 2025 that pushed its valuation past $1 billion.[2][3]

| Date | Round | Amount | Lead Investors | Valuation |
|---|---|---|---|---|
| June 2023 | Series A | $58 million | DST Global Partners, Radical Ventures | ~$300 million[2] |
| July 2025 | Growth Round | $110 million | Nvidia, Snowflake | ~$1 billion[3] |

Total disclosed funding: approximately $168 million.

Notable investors across rounds include DST Global Partners, Radical Ventures, Snowflake Ventures, Nvidia, and angel investor Nat Friedman (former GitHub CEO).[2] The 2024 Snowflake acquisition discussion, which would have valued Reka at more than $1 billion in cash, was the closest the company came to leaving the independent startup track.[5] After those talks ended without a deal in May 2024, the partners chose to keep working together on the commercial side and let Reka raise its own growth round in 2025 instead.[3]

## Offices and team

Reka is headquartered at 530 Lawrence Expy, PMB 9004, Sunnyvale, California. The company also operates Reka AI Pte. Ltd., a Singapore entity at 1 Robinson Road, AIA Tower, which handles operations across Asia. Reka AI Ltd., the UK entity, gives the company a presence in London. Beyond these formal offices, Reka describes itself as a remote-first company with talent based in California, Seattle, London, Zurich, and Hong Kong. As of mid-2025 the company had grown to roughly 50 people, rising to about 57 by April 2026, still small by frontier-lab standards but several times the size of the original 20-person research team that built Reka Core.

## Comparison with competitors

| Feature | Reka AI | [OpenAI](/wiki/openai) | [Google DeepMind](/wiki/google_deepmind) | [Anthropic](/wiki/anthropic) |
|---|---|---|---|---|
| Founded | 2022 | 2015 | 2010 (DeepMind) | 2021 |
| Headquarters | Sunnyvale, CA | San Francisco, CA | London, UK | San Francisco, CA |
| Team Size (approx.) | ~57 | ~3,000+ | ~2,500+ | ~1,500+ |
| Flagship Model (2024) | Reka Core | [GPT-4](/wiki/gpt-4) | [Gemini](/wiki/gemini) Ultra | [Claude](/wiki/claude) 3 Opus |
| Native Multimodal | Yes (text, image, video, audio) | Yes (GPT-4V/o) | Yes (Gemini) | Partial (text, image) |
| Video Understanding | Yes | Limited | Yes | No |
| Audio Input | Yes | Yes (Whisper integration) | Yes | No |
| On-Premise Deployment | Yes | No (API only) | Limited | No (API only) |
| Open-Source Models | Yes (Flash 3, Flash 3.1, Edge 2603, Reka Quant) | Limited (Whisper, CLIP) | Yes (Gemma) | No |
| Estimated Valuation | ~$1B (2025) | ~$157B (2025) | Part of Alphabet | ~$60B (2025) |

## ELI5: Reka AI in simple terms

Reka AI is a small company that builds smart computer programs (AI models) that can look at pictures and videos, listen to audio, and read text all at once, 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. They are proud that their tiny team built models almost as good as the giant labs' models for a lot less money. A big data company called Snowflake once tried to buy Reka for about a billion dollars, but the deal fell through, and instead Nvidia and Snowflake put money into Reka in 2025, which made it worth more than a billion dollars. Reka also gives away one of its models, Reka Flash 3, for free so anyone can use it.

## See also

- [Large Language Models](/wiki/large_language_model)
- [Multimodal AI](/wiki/multimodal_ai)
- [GPT-4](/wiki/gpt-4)
- [Gemini](/wiki/gemini)
- [Claude](/wiki/claude)
- [Transformer](/wiki/transformer)
- [Snowflake AI](/wiki/snowflake_ai)

## References

1. Reka Team. "Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models." arXiv preprint arXiv:2404.12387 (April 2024). https://arxiv.org/abs/2404.12387
2. Wiggers, Kyle. "Reka emerges from stealth to build custom AI models for the enterprise." TechCrunch (June 27, 2023). https://techcrunch.com/2023/06/27/reka-emerges-from-stealth-to-build-custom-ai-models-for-the-enterprise/
3. Bass, Dina. "Snowflake, Nvidia Back New Unicorn Reka AI in $110 Million Deal." Bloomberg (July 22, 2025). https://www.bloomberg.com/news/articles/2025-07-22/snowflake-nvidia-back-new-unicorn-reka-ai-in-110-million-deal
4. Wiggers, Kyle. "Snowflake teams up with Reka to add multimodal LLMs to data cloud." VentureBeat (2024). https://venturebeat.com/data-infrastructure/exclusive-snowflake-teams-up-with-reka-to-add-multimodal-llms-to-data-cloud
5. Bass, Dina. "Snowflake Talks to Acquire Reka AI Said to Fizzle With No Deal." Bloomberg (May 22, 2024). https://www.bloomberg.com/news/articles/2024-05-22/snowflake-talks-to-acquire-reka-ai-said-to-fizzle-with-no-deal
6. Tay, Yi. "Returning to Google DeepMind." Personal Blog (November 25, 2024). https://www.yitay.net/blog/returning-to-google-deepmind
7. Reka Team. "Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models." arXiv preprint arXiv:2405.02287 (May 2024). https://arxiv.org/abs/2405.02287
8. Reka AI. "Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms." Reka Blog (July 2025). https://reka.ai/news/reka-secures-110-million-to-accelerate-adoption-of-its-multimodal-ai-platforms
9. Wiggers, Kyle. "Reka launches Yasa-1, a multimodal AI assistant to take on [ChatGPT](/wiki/chatgpt)." VentureBeat (October 2023). https://venturebeat.com/ai/reka-launches-yasa-1-a-multimodal-ai-assistant-to-take-on-chatgpt
10. Reka AI. "[Reasoning](/wiki/reasoning) with Reka Flash 3." Reka Blog (March 2025). https://reka.ai/news/introducing-reka-flash
11. Reka AI. "Reka Flash 3.1 and Reka Quant." Reka Blog (2025). https://reka.ai/news/reka-flash-3-1-and-reka-quant
12. Dani Yogatama. Personal Homepage. https://dyogatama.github.io/
13. Mikel Artetxe. Personal Homepage. https://www.mikelartetxe.com/
14. Cyprien de Masson d'Autume. Crunchbase Profile. https://www.crunchbase.com/person/cyprien-de-masson-d-autume
15. Snowflake. "Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka." Snowflake Blog. https://www.snowflake.com/en/blog/multimodal-llm-snowflake-reka/
16. Reka AI. "Reka Core: Our Frontier Class Multimodal Language Model." Reka Blog (April 15, 2024). https://reka.ai/news/reka-core-our-frontier-class-multimodal-language-model
17. Oracle Corporation. "Oracle and Reka Collaborate to Advance AI Innovation." PR Newswire (April 18, 2024). https://www.prnewswire.com/news-releases/oracle-and-reka-collaborate-to-advance-ai-innovation-302120884.html
18. Reka AI. "Reka launches Nexus, an AI workforce powered by its state-of-the-art multimodal reasoning model." PR Newswire (March 10, 2025). https://www.prnewswire.com/news-releases/reka-launches-nexus-an-ai-workforce-powered-by-its-state-of-the-art-multimodal-reasoning-model-302396904.html
19. Reka AI. "Reka Vision: Intelligence Made Visible." Reka Blog (July 8, 2025). https://reka.ai/news/reka-vision-intelligence-made-visible
20. Reka AI. "Reka Quantization Technology." Reka Blog (July 10, 2025). https://reka.ai/news/reka-quantization-technology
21. Reka AI. "Reka Edge: Frontier-Level Edge Intelligence for Physical AI." Reka Blog (March 11, 2026). https://reka.ai/news/reka-edge-frontier-level-edge-intelligence-for-physical-ai
22. Reka AI. "Reka and Turing Partner to Pioneer Agentic Video Surveillance Platform." Reka Blog. https://reka.ai/news/reka-and-turing-partner-to-pioneer-agentic-video-surveillance-platform
23. Reka AI. "Using NVIDIA AI Blueprint for Video Search and Summarization with Reka Vision Agent." Reka Blog. https://reka.ai/news/using-nvidia-ai-blueprint-for-video-search-and-summarization-with-reka-vision-agent
24. Hugging Face. "RekaAI/reka-flash-3." https://huggingface.co/RekaAI/reka-flash-3
25. Hugging Face. "RekaAI/reka-flash-3.1-rekaquant-q3_k_s." https://huggingface.co/RekaAI/reka-flash-3.1-rekaquant-q3_k_s
26. Hugging Face. "RekaAI/reka-edge-2603." https://huggingface.co/RekaAI/reka-edge-2603
27. GitHub. "reka-ai/rekaquant." https://github.com/reka-ai/rekaquant
28. TipRanks. "Reka AI Showcases Multimodal Security Platform at Smart City Asia 2026." (May 2026). https://www.tipranks.com/news/private-companies/reka-ai-showcases-multimodal-security-platform-at-smart-city-asia-2026
29. Reka AI. "Reka and Moonvalley Join Forces to Advance Models and Infrastructure for Physical AI." PR Newswire (June 11, 2026). https://www.prnewswire.com/news-releases/reka-and-moonvalley-join-forces-to-advance-models-and-infrastructure-for-physical-ai-302797869.html