Reka

Reka is an AI company founded in 2022 by researchers from Google DeepMind, Meta, and Baidu, led by CEO Dani Yogatama and Chief Scientist Yi Tay, who previously led development of PaLM and UL2 at Google. The company builds a family of multimodal language models trained from scratch, meaning every model in its lineup handles text, images, video, and audio as native first-class inputs, not through post-hoc adapters. Reka's core proposition is bringing frontier-class multimodal reasoning to enterprises that need deployment flexibility: cloud, on-premises, VPC-isolated, or fully air-gapped.

The current lineup spans four model tiers. Reka Core (67B) is the flagship model, handling complex reasoning tasks across all four input modalities with a 128,000-token context window. Reka Flash (21B) is the cost-performance tier: the Flash 3 variant was open-sourced in March 2025 under Apache 2.0, making it available for self-hosted deployments. Reka Edge (7B) targets real-time inference on constrained hardware, and Reka Spark (1B) is designed for on-device embedding. All models are accessed via a RESTful API with Python and JavaScript SDKs, and the company's Reka Vision platform handles large-scale video and image search for enterprise customers including Shutterstock and Turing Video.

Reka at a glance, April 2026

Reka's model family currently comprises four production models. Reka Core is the 67B flagship, training from scratch on 5 trillion deduplicated tokens across 32 languages with a 128K context window. Reka Flash 21B was the first model to earn significant community attention when it matched Gemini Pro on MMLU and GPQA benchmarks despite being trained by a team of roughly 20 researchers. Reka Flash 3, open-sourced in March 2025 under Apache 2.0, competes with QwQ-32B at 34% fewer parameters and supports 32K context. Flash 3.1 followed in July 2025 with reinforcement learning improvements. Reka Edge (7B) is the fastest in-class vision-language model at its parameter count, outperforming Gemma 7B and Mistral 7B on multimodal tasks. Reka Spark (1B) targets sub-device embedded deployments.

On the products side, Reka offers three access surfaces: the Chat API for LLM-style interactions with multimodal inputs, the Research API (reka-flash-research) for agentic multi-step reasoning tasks with parallel thinking modes, and Reka Vision for video and image search at scale. Deployment modes include API, managed cloud, on-prem, VPC, and air-gapped, giving Reka a deployment flexibility profile closer to enterprise database vendors than to OpenAI or Google.

What Reka is actually good at

Reka's genuine edge is native video understanding. When the Core tech report was published in April 2024, Reka Core scored 59.3% on the Perception Test video benchmark, ahead of Gemini Ultra at 58.7%. GPT-4o does not process continuous video clips natively at all, handling only extracted frames. For enterprises that need to query large video archives, index surveillance footage, or run inference over media libraries, Reka Core and Reka Vision are purpose-built in a way that general-purpose frontier models are not.

Reka Flash 3's open-source release is the other strong card. Developers who want a capable 21B reasoning model they can run on-prem, self-host, fine-tune, and modify without licensing restrictions now have an Apache 2.0 option that benchmarks comparably to proprietary models in its weight class. The MMLU-Pro score of 65.0 and strong AIME-2024 math reasoning put it in credible territory for coding, instruction following, and function-calling tasks.

Multilinguality is also a real differentiator. With 32-language support and strong performance on translation benchmarks (83.2 on COMET WMT'23 for Flash 3), Reka competes well in deployments where European and Asian language coverage matters alongside vision tasks.

"We are excited to partner with Reka to bring Reka Core's impressive industry leading performance to customers through Snowflake Cortex." - Baris Gultekin, Head of AI at Snowflake, Reka press release, April 15, 2024

"Reka's dynamic multimodal and multilingual models allow enterprises to unlock more value from their data." - Greg Pavlik, SVP AI and Data Management at Oracle, Reka press release, April 15, 2024

Where Reka breaks: the failure modes users keep hitting

The playground video cap is the first frustration that surfaces in practitioner discussions. The chat.reka.ai interface restricts video uploads to 30MB and one minute. For media professionals or enterprise teams testing with real-world footage, this is a hard wall that forces a switch to the API before doing any meaningful evaluation.

Enterprise customization (fine-tuning, on-prem deployment, VPC isolation) is not self-serve. There is no transparent pricing for anything beyond standard API token costs. Teams that need custom deployment have to contact sales, sit through demos, and negotiate a contract before they can assess whether Reka actually fits their workload, creating months-long friction compared to Google Vertex AI or AWS Bedrock.

Ecosystem depth is thin relative to OpenAI and Google. Community-built tutorials, third-party integrations, LangChain plugins, and public benchmarks referencing Reka are sparse. A developer picking a model for a greenfield project will find far more pre-built tooling for GPT-4o or Gemini than for any Reka model. Reka Flash 3 has begun to change this for the open-source community, but the proprietary-tier models remain under-documented outside official channels.

Knowledge cutoff is another gap. Reka Core's training data knowledge cutoff is November 2023, which was already dated at launch in April 2024 and is increasingly stale as of 2026. OpenAI and Google both refresh model knowledge cutoffs more aggressively.

Finally, the Research API's parallel thinking modes carry a steep cost multiplier: $35 per thousand requests for "low" parallel thinking versus $25 standard, and $60 per thousand for "high" parallel thinking. Teams that explore agentic workflows casually can run up unexpectedly large bills.

Reka vs. GPT-4o vs. Gemini 1.5 Pro

Architecture and video handling. GPT-4o is an omnimodal transformer estimated at roughly 200B parameters. It handles text, images, and audio natively, but video processing in GPT-4o operates on extracted frames rather than continuous video streams. Reka Core processes full video clips as a continuous modality, which is a meaningful architectural difference for surveillance, media, or scientific video analysis tasks. Reka Core's context window is 128K tokens, matching GPT-4o. Pricing at launch was comparable ($10/$25 per 1M tokens input/output), but Reka has since dropped Core pricing to $2.00/$6.00 per 1M tokens, below GPT-4o's $2.50/$10.00. OpenAI does not offer on-premises deployment; Reka does.

Context window and deployment. Gemini 1.5 Pro holds the architectural advantage that matters most for long-document enterprise work: a 1 million token context window, roughly 8x Reka Core's 128K. Gemini 1.5 Pro also accepts native video input with clips up to one hour, compared to Reka's playground cap of one minute and the API's multi-minute ceiling. Cost is where Gemini pulls further ahead for casual use: Gemini 1.5 Flash at $0.075/$0.30 per 1M tokens is dramatically cheaper than Reka Flash at $0.80/$2.00. However, Gemini is exclusively available via Google Cloud (Vertex AI) and has no on-prem path. Reka supports air-gapped enterprise deployment that Google cannot match. On the April 2024 video perception benchmark, Reka Core marginally edged Gemini Ultra (59.3% vs 58.7%), though Gemini's later models have continued advancing.

Open weights. Neither GPT-4o nor Gemini 1.5 Pro release weights publicly. Reka Flash 3 (21B, Apache 2.0) is downloadable from Hugging Face and can be self-hosted, fine-tuned, and modified. This is a concrete differentiation for organizations with data sovereignty requirements or teams that cannot route sensitive data through third-party APIs.

Is the paid tier worth it?

Reka's per-token pricing is honest and pay-as-you-go with no upfront commitment. New accounts receive $20 in free credits, enough for meaningful testing. For standard text workloads, Reka Flash at $0.80/$2.00 per 1M tokens is competitive with mid-tier models, though Gemini 1.5 Flash undercuts it significantly. Reka Core at $2.00/$6.00 per 1M tokens is cheaper than Claude 3 Opus and comparable to GPT-4o, but without the ecosystem depth.

Where the economics favor Reka is in three scenarios: (1) video inference, where Reka Core at $0.08 per video-minute is priced for production-scale use; (2) on-prem deployments where Reka Flash 3 as open-weight Apache 2.0 eliminates per-token API costs entirely for self-hosted teams; and (3) multilingual enterprise workloads in Asian or European languages where Reka's 32-language training pays off.

The Research API's pricing structure ($25-60 per thousand requests depending on thinking mode) is better understood as an agentic pipeline cost than a chat API cost. For single queries it's expensive; for replacing a human research analyst workflow, the math changes. Enterprise pricing for custom deployments requires a sales call and is not public.

Best use cases (and when to skip it)

Use Reka when: Your application depends on native video understanding at scale, you need multimodal inference with on-premises or air-gapped deployment, you want an Apache 2.0 21B reasoning model you can self-host without vendor lock-in, or you are building multilingual applications that need strong Asian and European language coverage alongside vision capabilities. Media companies, enterprise security teams, healthcare organizations processing imaging alongside patient records, and mining or manufacturing companies with video-heavy sensor data are Reka's natural customers.

Skip Reka when: You need a 1M+ token context window for long-document work (use Gemini 1.5 Pro). You need the broadest ecosystem of third-party integrations and tutorials (use GPT-4o). You are an individual user wanting a conversational assistant for everyday tasks (ChatGPT Plus is better suited). Your team lacks the engineering resources to integrate a raw API, or your budget requires staying in a vertically integrated platform (Google AI Studio or OpenAI's ChatGPT Enterprise are simpler starting points). Cost-sensitive text-only workloads may find Gemini Flash or Llama-based alternatives cheaper.

Getting started with Reka

Reka's API is available at docs.reka.ai with Python and JavaScript SDKs. New accounts get $20 in credits, enough to test Core, Flash, and Spark across text and image inputs before committing spend. The Reka Playground at chat.reka.ai offers browser-based access to test multimodal inputs (note the 30MB/1-minute video cap in the playground). For the open-weight path, Reka Flash 3 and Flash 3.1 are available on Hugging Face under Apache 2.0. Model weights download at 39GB (fp16) or 11GB at 4-bit quantization, making Flash 3 runnable on a single A100 GPU. Enterprise deployments on-prem or in VPC require contacting Reka's sales team via reka.ai.