Curiosity · AI Model
Reka Core
Reka Core is the flagship multimodal model from Reka AI, a San Francisco lab with deep research roots (ex-DeepMind, Meta, Google Brain). Launched in April 2024, it was an unusually ambitious debut — combining text, image, audio, and video understanding with a 128k context window, and positioning Reka as a frontier-tier player alongside OpenAI, Anthropic, and Google.
Model specs
- Vendor
- Reka AI
- Family
- Reka
- Released
- 2024-04
- Context window
- 128,000 tokens
- Modalities
- text, vision, audio, video
- Input price
- $3/M tok
- Output price
- $15/M tok
- Pricing as of
- 2026-04-20
Strengths
- Native support for video and audio in one endpoint
- 128k context alongside full multimodality
- Enterprise-friendly deployment options (VPC, private cloud)
Limitations
- Smaller ecosystem than OpenAI / Google / Anthropic
- Superseded on some tasks by GPT-4o and Gemini 2.5 Pro
- Less public third-party benchmarking than peer frontier models
Use cases
- Multimodal agents that need image, audio, and video in one model
- Enterprise-hosted alternatives to GPT-4o / Gemini
- Research on non-big-three frontier models
- Video question answering over short clips
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| MMMU | ~56% | 2026-04 |
| MMLU | ~83% | 2026-04 |
| Perception Test | ~60% | 2026-04 |
Frequently asked questions
What is Reka Core?
Reka Core is Reka AI's flagship multimodal frontier model — able to process text, images, audio, and video with a 128k context window. It launched in April 2024 as Reka's most capable public model.
Who builds Reka Core?
Reka AI is a San Francisco lab founded in 2022 by former DeepMind, Meta, and Google Brain researchers. The team ships frontier multimodal models with a focus on enterprise deployments.
Sources
- Reka AI homepage — accessed 2026-04-20
- Reka Core technical report (arXiv) — accessed 2026-04-20