Reka Flash Efficient And Capable Multimodal Language Models
Source
reka.ai
Author
Reka AI
Date
Context
Reka AI's February 2024 Flash announcement described a 21-billion-parameter model evaluated on language tasks, questions about images and video, and video captioning. Its smaller, seven-billion-parameter Edge variant targeted local deployment and applications where response time matters.
The release separated two capabilities that can easily be conflated: Flash was pretrained on text from over 30 languages, but the deployed models' image and video interactions worked best in English at launch. A multilingual text model therefore did not imply equally strong visual assistance in every supported language.
Terms in this piece · Glossary
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
RLHF — Reinforcement learning from human feedback — training a model to prefer answers humans rate as better, which turns a raw text predictor into a usable assistant.
