Our models are trained from scratch with our proprietary technology to support multimodal reasoning. Beyond chat bots, they are designed to become powerful foundations for action bots that can accomplish tasks for you.

Model Architecture

Our models are based on a novel multimodal transformer architecture that can take as input interleaved images, audio, video, and text and output text or audio tokens.

Complex Reasoning

Our models have excellent reasoning and general chat capabilities, including logical, physical, and temporal reasoning.

Visual Understanding

Our models understand image and video with arbitrary resolution and aspect ratio. They can perform OCR, understand visual PDFs, tables, charts, and diagrams. For videos, our models understand up to 5 minute clips and can process longer or live videos by streaming them.

Audio Understanding

Our models understand multilingual audio natively without using a separate speech recognition.

Instruction Following and Chaining

Our models are trained to follow complex multi-step instructions, making it suitable as a building block to support agentic tasks.

Capabilities

Coding, Function Calling, and Tool Use

Our models are proficient coders with supports for tool use via function calling and code execution, allowing them to perform actions to accomplish tasks going beyond a regular chatbot.

Speech Output

Our models support generating audio tokens that can be decoded into voice.

Model Specifications

  • Reka Spark

    2B parameters

    Fully multimodal

    Multilingual in 32 languages

    128K context length

    Ideal for deployment on smartphones, robots, wearable devices, and home appliances

  • Reka Edge

    7B parameters

    Fully multimodal

    Multilingual in 32 languages

    128K context length

    Ideal for deployment on laptops, desktops, or high-end tablets.

  • Reka Flash

    21B parameters

    Fully multimodal

    Multilingual in 32 languages

    128K context length

    Ideal for deployment on premises or in a private cloud.

  • Reka Core

    67B parameters

    Fully multimodal

    Multilingual in 32 languages

    128K context length

    Ideal for most complex use cases, model distillation.

Benchmark Results

Use Our Models Today

  • Reka Chat

    Interact with our models via Reka Chat.

  • Build with our API

    You can also build your own applications via our API.