> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/reka.md).

# Reka

#### Reka: Multimodal AI Models for Enterprise and Research

**What is Reka?**

Founded in 2022 by experts from Google DeepMind, Meta FAIR, and Baidu, Reka is a multimodal AI company building models that natively understand and reason across text, images, video, and audio. Co-founders Dani Yogatama and Chief Scientist Yi Tay lead the team. Reka secured $110 million in funding from NVIDIA and Snowflake, reaching unicorn status at over $1 billion valuation.

***

**Core Model Lineup**

**Reka Edge (7B)** — Dense, 7B parameter model designed for resource-constrained environments. State-of-the-art for its compute class, often surpassing much larger models on benchmarks.

**Reka Flash (21B)** — One of the few publicly available models supporting interleaved multimodal inputs — text, image, video, and audio — in a single model with a 128K token context window. Delivers outsized performance at 21B parameters, making it a cost-effective solution for enterprise multimodal applications.

**Reka Core** — The largest and most capable model. At release, it approached frontier models from OpenAI, Google, and Anthropic on both automatic and blind human evaluations. On multimodal chat, Core ranked as the second most preferred model under a blind third-party evaluation, outperforming Claude 3 Opus. On video question answering via Perception-Test, Core outperforms Gemini Ultra. On image question answering, it is competitive with GPT-4V.

***

**Reka Flash 3 (March 2025)**

A general-purpose reasoning model released as open-source, with 21B parameters. Competitive with OpenAI's o1-mini on reasoning benchmarks. Trained from scratch on diverse public and synthetic datasets, with supervised fine-tuning and REINFORCE Leave-One-Out (RLOO) reinforcement learning.

Key features include a 32K token context window, a budget forcing mechanism using reasoning tags that allows developers to cap the model's thinking steps for latency control, and 4-bit quantization support compressing the model from 39GB to just 11GB without significant performance loss. Released in Llama-compatible format, compatible with any Llama-supporting library.

***

**Reka Flash 3.1 (July 2025)**

An updated version of Flash 3 with significant improvements in coding and agentic task performance. The improvement comes from a new RL algorithm and enhanced infrastructure enabling training at much larger scale. It improves by 10 points on LiveCodeBench v5 from Flash 3, and is competitive with Qwen3-32B, o3-mini, and Gemini 2.5 Flash Thinking on coding benchmarks.

A multimodal version of Flash 3.1 serves as the base for Reka's two flagship products: Reka Research (a web and document-browsing agentic AI) and Reka Vision (a visual understanding and search platform). Both models are freely available on Hugging Face and GitHub.

***

**Multimodal Capabilities**

Reka Flash supports arbitrary image resolution and aspect ratio with enhanced OCR and document, table, and chart understanding. For video, it provides high-level temporal understanding with native audio capture, supporting videos of 3 to 5 minutes with even longer content processable via streaming. A Reka Quant quantization library accompanies Flash 3.1 to support local deployment.

***

**Enterprise and Government**

Reka's solutions are available through US government procurement contracts including SEWP V, ITES-SW2, and NASPO ValuePoint, via a partnership with Carahsoft. Use cases include public safety video search, live event detection, person tracking across surveillance systems, multilingual access, and compliance-driven intelligent data use.

***

**Strengths**

* Native multimodal architecture: text, image, video, audio in a single model
* Strong benchmark performance at small parameter counts relative to model size
* Open weights for Flash 3 and Flash 3.1 freely available
* Budget forcing mechanism for precise inference cost control
* Llama-compatible format for easy integration
* Backed by NVIDIA and Snowflake with strong enterprise partnerships

**Limitations**

* Core model is proprietary and not openly available
* Flash 3 is primarily English-only
* Smaller team and research footprint than major labs
* Less ecosystem tooling compared to models from Databricks, NVIDIA, or IBM