Overview
Reka AI is a multimodal AI company founded by former DeepMind, Meta, and Google Brain researchers. Their unique multimodal architecture handles text, code, images, video, and audio natively in a single model. In May 2026 Reka merged with video-generation company Moonvalley in an all-share deal, bringing in researchers from DeepMind, Meta, Google, Wayve, and Runway, and has since shifted its center of gravity toward "physical AI" — world models and robotics. The model line spans four sizes: Reka Core (frontier-class), Reka Flash 3 (efficient mid-tier), Reka Edge, and Reka Spark.
Lineup
| Model | Params | Use Case |
|---|---|---|
| Reka Core | 67B | Frontier-class multimodal |
| Reka Flash 3 | 21B | Efficient, general-purpose |
| Reka Edge | 7B | Edge devices |
| Reka Spark | 2B | Ultra-lightweight |
Reka Core (frontier class)
A 67B multimodal model handling text, images, video, and audio natively. It places near the top on image QA (MMMU, VQAv2) and in blind human evaluation of multimodal chat.
Reka Flash 3 (2026 Latest)
21B general-purpose, instruction-tuned model.
| Metric | Value |
|---|---|
| Params | 21B |
| Context | 32K |
| Optimization | Reinforcement learning |
| Quantization | 11GB at 4-bit |
| Reasoning | Explicit thinking tags |
- Excels at chat, coding, instruction following, function calling
- ELO score improved by 43 points (internal)
- Overall quality: 66.1% → 72.2%
Key Features
1. Native Multimodal
Handles text, images, video, and audio in a unified model. Supports temporal video understanding and environmental sound recognition.
2. OCR & Document Understanding
Strong at documents, tables, charts, and diagrams.
3. Function Calling
Structured output for function calling.
4. Efficient Quantization
Flash 3 fits in 11GB at 4-bit — runs on standard GPUs.
5. Explicit Reasoning Tags
Flash 3 visualizes its thought process with explicit reasoning tags.
Products
Beyond the models themselves, Reka ships a set of business-facing products.
| Product | What it does |
|---|---|
| Reka Nexus | Enterprise platform for creating and managing AI workers on Reka Flash (21B) to automate workflows |
| Reka Vision | Large-scale video and image analysis with highlight extraction |
| Reka Research | Agentic web research |
Access and pricing
API rates run from $0.10 to $2.00 per 1M input tokens, with Reka Flash 3 the cheapest at $0.10 per 1M. Reka is also available through Snowflake Cortex and as an on-premises deployment.
Strengths
- ✅ Native multimodal (including video and audio)
- ✅ Multiple sizes (67B/21B/7B/2B)
- ✅ Reka Nexus / Vision / Research round out the model line with business products
- ✅ Function calling and structured output
- ✅ Efficient quantization (11GB for Flash 3)
- ✅ Native video/audio understanding
Weaknesses
- ❌ Low brand awareness
- ❌ Limited developer tools
- ❌ Small ecosystem
- ❌ Limited Japanese support
- ❌ Flash 3 context is 32K (shorter than competitors)
Official Resources
- Website: https://reka.ai
- Reka Playground: https://chat.reka.ai
- API Docs: https://docs.reka.ai
- Tech Report: https://publications.reka.ai