EXPLORINEER
en

Archive — US — turtlesai

Reka Flash 3: a 21B reasoning model with great performance

Reka has open-sourced Reka Flash 3, a 21 billion-parameter multimodal language model designed for tasks like chatting and function calls. It excels in performance compared to proprietary solutions, with support for contexts up to 32,000 tokens, making it suitable for low-latency applications. Despite its efficiency, it has limitations in handling extensive knowledge tasks.

Key points

  • Reka Flash 3 is a 21 billion parameter open-source model.
  • It shows competitive performance with proprietary models such as OpenAI's o1-mini.
  • Supports context lengths up to 32,000 tokens, ideal for various applications.
  • Optimized for low latency and local device deployments.
  • Training involved pre-training on diverse data, tuning with high-quality instructions, and reinforcement learning.
  • Outperformed previous models on AIME-2024 and achieved a COMET score of 83.2 on WMT'23.
  • Scored 65.0 on MMLU-Pro, indicating it may need to integrate with web search for knowledge-intensive tasks.
  • Features a unique mechanism for reasoning management using tags, controlling reasoning steps during output generation.
  • Compressed model size can go down to 11 GB while maintaining high performance, a significant efficiency compared to larger models.
  • Designed primarily for English but shows some multilingual capabilities; alignment and training remain areas for future improvement.
  • A trial version and downloadable model weights are available under Apache 2.0 license for developers.

Why it matters

  • The release of Reka Flash 3 aligns with ongoing trends in AI toward open-source and accessible models.
  • As of 2025, low-latency applications are crucial for AI deployment, especially in edge computing.
  • The model's multi-tasking capabilities reflect the industry's shift towards versatile language models that serve various applications.
  • Open-source projects significantly influence the pace of AI development and access to advanced tools for researchers.

Reka Flash 3 represents a major advancement in multimodal language models, combining innovative features with efficiency, inviting further exploration and potential applications in the AI landscape.