Reka Flash 3

Reka Flash 3 is a general-purpose, instruction-tuned large language model from Reka AI, a provider that builds its own family of models. It carries 21 billion parameters and was released by the provider on 12 March 2025. The model sits in the text category, so it is aimed at language work rather than image, audio or video processing.

On the capability side, Reka Flash 3 is tuned for general chat, coding tasks, instruction-following and function calling. It accepts text as input and returns text as output. The context window is 65,536 tokens, and the maximum output length is 58,982 tokens, which means a single request can carry a large prompt and still leave room for a long completion. Its knowledge cutoff is 31 January 2025, so anything after that date is outside what the model was trained on. Reasoning output is supported, which lets you surface the model's intermediate thinking when you want it.

Those traits point to a fairly broad set of uses. Chat assistants and instruction-following agents are a natural fit, since the model is instruction-tuned and handles multi-turn conversation. Coding help, such as generating, explaining or refactoring code, follows from the coding focus. Function calling makes it suitable for tool-using workflows, where the model decides which function to invoke and with what arguments. The long context window helps with tasks that need a lot of material in one request, for example summarising a lengthy document or answering questions over a large body of text.

You call the model through api14's OpenAI-compatible API using the model id "rekaai/reka-flash-3". Because the interface follows the OpenAI chat completions shape, you can point an existing OpenAI client at api14 and change the model string. The model supports these request parameters: frequency_penalty, include_reasoning, logprobs, max_tokens, presence_penalty, reasoning, seed, stop, structured_outputs, temperature, top_k, top_logprobs and top_p. In practice that covers sampling controls (temperature, top_p, top_k), repetition controls (frequency_penalty, presence_penalty), output length (max_tokens), deterministic sampling (seed), stopping conditions (stop), probability inspection (logprobs, top_logprobs), reasoning control (reasoning, include_reasoning) and structured_outputs for constraining the response format.

One thing to keep in mind is scope. Reka Flash 3 is text-only, so if your application needs to process images, audio or video directly, this model is not the right choice. The 65,536-token context window is generous but finite, and very large inputs will need chunking or retrieval rather than a single pass. The knowledge cutoff of 31 January 2025 also means the model will not know about events after that date unless you supply the information in the prompt.