Reka Flash 3

What it is

Reka Flash 3 is a general-purpose reasoning model with 21 billion parameters, trained from scratch rather than adapted from an existing base model ^[1]. Its weights were uploaded to Hugging Face on 11 March 2025 ^[1], and the release is under the Apache 2.0 licence ^[1]. In this context, a reasoning model is one that writes out intermediate steps before giving a final answer. Reka Flash 3 puts that thinking inside `<reasoning>` tags, so a program reading the output can tell the deliberation apart from the answer itself ^[1].

How it was trained

Training ran in two stages ^[1]. The first was supervised fine-tuning on a mix of synthetic and public datasets ^[1]. The second applied RLOO, a reinforcement-learning method that compares several answers sampled for the same prompt and pushes the model toward the better ones, using both model-based and rule-based rewards ^[1]. Model-based rewards mean another model does the judging; rule-based rewards mean a program checks something mechanical, such as whether a required format was followed or a checkable answer is right ^[1]. The model card lists both kinds as part of the recipe ^[1].

Prompting and the reasoning block

The model uses the cl100k_base tokenizer and adds no special tokens of its own ^[1]. Prompts follow a human-and-assistant format, with the two parts separated by a `<sep>` token, and generation should stop when `<sep>` appears ^[1]. Thinking is wrapped in `<reasoning>` tags ^[1]. Reka also describes a budget-forcing trick: the thinking block can be cut off after a chosen number of steps, and the model still returns a usable answer ^[1]. That gives whoever runs the model a way to cap how long it deliberates before it has to respond, which matters when answers have to come back within a fixed time or token cost ^[1].

Packaging and use

The weights ship in a Llama-compatible format, so existing Llama tooling can load them without a new loader being written ^[1]. The model also powers Nexus, the company's platform for building and managing AI workers ^[1]. Open weights under a permissive licence mean the model can be downloaded, inspected and run on the user's own hardware, instead of only being reachable through an API.

Limits

The model card calls Reka Flash 3 an English-only model, and notes that it often thinks in English even when the prompt is in another language ^[1]. For knowledge-intensive tasks, Reka recommends pairing it with web search rather than relying on what the model stored during training ^[1].

Where it fits

Reka says the model performs competitively with OpenAI's o1-mini ^[1], a comparison that places it among the smaller reasoning models rather than the largest. Its distinguishing features are the combination of a from-scratch training run, open weights under Apache 2.0, and packaging that slots into tooling already built around Llama checkpoints ^[1]. For people who want to study or build on a reasoning model directly — reading its thinking traces, controlling their length, or running it locally — those choices are what make the release useful.