Reka-Flash-21B-online

Reka Flash-21B is a multimodal and multilingual language model developed by Reka AI that offers efficient performance rivaling larger models. It excels in various language and vision benchmarks, multilingual reasoning, and visual and video processing.

Key features:
- Evaluated on benchmarks like MMLU, GSM8K, HumanEval, and GPQA, outperforming models like Llama 2 70B and GPT-3.5
- Pretrained on text from over 32 languages, excelling in multilingual commonsense reasoning, causal reasoning, and question answering
- Competitive performance in visual question answering, video captioning, and video question answering
- Instruction-tuned and fine-tuned with reinforcement learning from human feedback using Proximal Policy Optimization
- Reka Edge, a compact 7B variant, outperforms similar-scale models like Mistral 7B and Llama-2 7B on standard language benchmarks

How it works:
Users can interact with Reka Flash-21B through the Reka Python SDK or HTTP API. After obtaining an API key from the Reka Platform Dashboard, users can install the SDK and start querying the models.

Integrations:
- Python SDK for developers to integrate Reka models into their applications
- HTTP API for flexible interaction with the models from different programming environments

Use of AI:
Reka Flash-21B is trained from scratch and optimized to push its performance limits. It uses advanced techniques like reinforcement learning from human feedback with Proximal Policy Optimization for fine-tuning.

AI foundation model:
The model is rigorously evaluated on multiple benchmarks to ensure high performance across various tasks.

How to access:
- Available in the Reka Playground for public beta testing
- Accessible through the Python SDK and HTTP API for integration into various applications