Reka AI
Multimodal models and video infrastructure for physical AI
Overview
Reka is a Sunnyvale-based AI lab that now describes itself as building models and infrastructure for physical AI. It sells an OpenAI-compatible API serving its Reka Edge and Flash 3 models plus hosted open models, and Reka Vision for searching, tagging and clipping video libraries. It merged with video-generation startup Moonvalley in June 2026.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Text, image and video input
- 16k context
- Tools and structured outputs
- Weights on Hugging Face under BSL 1.1
- 21B reasoning model
- 64k context, text input
- Structured outputs
- Apache 2.0 weights
- DeepSeek V4 Flash and Pro
- GLM 5.3 and GLM 5.3 Flash
- Qwen3.8 27B
- Cached-input rates where supported
- Indexing, search and Q&A
- For evaluating the Vision API
- Video search $0.005 per search
- Q&A and tagging $2 per 1M output tokens
- Clips $0.04 per input minute
- Videos auto-deleted after 30 days
- No rate limits
- Bulk discounts
- Flexible storage
- Dedicated support
Use cases
What you can produce with Reka AI
- Call Reka Edge, Reka Flash 3 or hosted DeepSeek, GLM and Qwen models through one OpenAI-compatible endpoint by changing a model ID.
- Send images and short videos alongside text in a chat request and get answers about what they show.
- Upload a video library to Reka Vision and find timestamped moments with a plain-language search.
- Ask questions across one or more videos and read back transcripts, captions, scenes and detected objects.
- Cut highlight clips from long footage automatically, including from new uploads on a YouTube channel.
- Run the Reka Edge weights on an Apple Silicon Mac or a single NVIDIA GPU with the vllm-reka plugin.
- Connect an AI agent to a video library through the Reka Vision MCP server.
Reka's most useful product today is Reka Vision: upload a video library, then search it in plain language, ask questions across footage, tag it and cut highlight clips, with three free hours before per-minute billing. Reka Edge 2603 is a capable small vision-language model you can call for $0.10 per million tokens or run locally, though its licence needs a commercial agreement once revenue reaches $1M.
The model line-up has shrunk to Edge and Flash 3, with the API now padded out by hosted DeepSeek, GLM and Qwen models, and the company's attention has moved to robotics and world models. Pick it for video understanding or an on-device multimodal model, not for frontier text reasoning.
What we found
Reka began in 2022 as a lab building natively multimodal models, staffed by researchers from DeepMind, Google Brain and Meta, and through 2024 it was known for the Core, Flash and Edge family. The company has since repositioned itself. Its homepage now calls Reka an AI research lab building models and infrastructure for physical AI, meaning robots, wearables and cameras that have to understand the world in real time. In June 2026 it merged with Moonvalley, the video-generation startup, bringing in Mateusz Malinowski and Mikołaj Bińkowski, both former Google DeepMind researchers who worked on models that became Veo. Bessemer's portfolio page describes the deal as Reka taking over Moonvalley. In October Reka released a research preview of Rho-1, a 19-billion-parameter model that reads and generates text, images and video and can output robotic actions.
The money behind it is real. Bloomberg reported in July 2025 that a 110 million dollar round backed by Nvidia and Snowflake took the valuation past 1 billion dollars, after 2024 acquisition talks with Snowflake ended without a deal.
What you can buy has narrowed and widened at once. Reka Core no longer appears in the API, and the Reka Research agent API has dropped out of the documentation. The API now serves two Reka models, Edge 2603 and Flash 3, alongside hosted open models from DeepSeek, Zhipu's GLM and Qwen, all through one OpenAI-compatible endpoint billed per token with no platform fee. Reka Vision, the strongest product, indexes video libraries for natural-language search, question-answering, tagging and clip generation, with three free hours of indexed video before per-minute billing.
Edge's weights are downloadable but under a source-available licence that requires a commercial licence once revenue reaches 1 million dollars; Flash 3 is Apache 2.0. Reka suits teams that need video understanding or a small multimodal model they can run themselves. Anyone wanting a general chatbot should look elsewhere.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
About this listing
This entry was compiled from publicly available data including Reka AI's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Reka AI unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Reka AI directly →
Spotted something out of date? Suggest an update →
Alternatives to Reka AI
More in LLMs & Chat
