Reka AI provides a suite of natively multimodal foundation models designed to understand and reason across text, images, videos, and audio. This "Multimodal-from-Scratch" approach allows for high-fidelity understanding with low latency, setting it apart from systems adapted from text-only origins. The company targets enterprise customers across diverse sectors, including media, security, defense, finance, and healthcare, offering solutions for complex data analysis and automation.
Native Multimodal Processing Without Text-Only Bottlenecks
Reka’s models are trained from inception on diverse data types, not adapted from text-only systems. This native multimodality enables them to reason about dynamic, real-world environments more effectively. For instance, Reka Clip uses these models for automated video clipping, smart captioning, and one-click content repurposing for social media, tasks that typically demand significant manual effort. This integrated processing of multimodal inputs enhances overall accuracy and contextual understanding.
Core, Flash, and Edge Models for Different Workloads
Reka offers a tiered family of models, each optimized for specific performance, cost, and deployment needs:
- Spark (1B parameters): Designed for edge devices.
- Edge (7B parameters): Ideal for fast vision-language tasks and real-time clip generation.
- Flash (21B parameters): Provides a balance of performance and cost.
- Core (67B parameters): The flagship model, built for complex multimodal tasks and competitive with models like GPT-4V and Claude-3 Opus.
These models support a large context window of up to 128K tokens, extendable to 400K tokens, accommodating approximately 100K English words, 100 images, and one hour of video. They’re also multilingual, pretrained on 32 languages with a focus on 12 major global languages.
On-Premise, Cloud, and API Deployment Options
Reka AI’s models power various enterprise applications:
- Media: Automating content creation, generating video highlights, and repurposing content.
- Security & Defense: Intelligent surveillance and threat detection.
- Finance & Healthcare: Streamlining operations and enhancing data analysis.
- E-commerce: Visual cataloging.
- Cybersecurity: Real-time analysis.
Deployment options are flexible, including API access, on-premises, on-device, VPC, and air-gapped environments. Reka also integrates with platforms such as Snowflake and Oracle.
Token-Based API Pricing with Enterprise Custom Quotes
Reka AI provides a free playground for users to test model capabilities. For broader use, the pricing structure varies:
| Plan | Price | Key Details |
|---|---|---|
| API Access | Consumption-based | Varies by model and product |
| Reka Chat | Token-based | Charges for input/output tokens and multimodal data usage, leading to variable costs |
| Reka Research | Flat rate per 1,000 requests | Offers more predictable budgeting |
| Enterprise Deployments | Negotiated contracts | Includes on-premise, VPC, and air-gapped options |
Limited Brand Recognition and Smaller Ecosystem
Reka AI has aspects that warrant consideration:
- Reka’s models require significant technical expertise to integrate and maintain in production environments.
- Like other AI systems, outputs can include hallucinations or inaccurate information, requiring human review.
- API pricing is usage-based, which can become costly for high-volume applications without careful monitoring.