Daraj
← All articles

Introducing Daraj-270M.

A compact 270M-parameter language model purpose-built to orchestrate real payment and ticketing tools on the DarajPay platform.

Daraj · November 16, 2026 · 7 min read

Daraj-270M - Open-source small language model built in Africa

We built Daraj-270M to prove a simple idea: a small, focused model can reliably operate real tools, in real workflows, at a fraction of the cost of a frontier model.

What is Daraj-270M

Daraj-270M is a 270-million-parameter causal language model fine-tuned from Gemma 3 270M Instruct. It is not a general-purpose chatbot. It is a tool-calling agent built to operate the DarajPay platform's Model Context Protocol (MCP) tools, handling account registration, catalog searches, ticket booking, and payment management.

Instead of generating free-form text, the model emits structured JSON containing a reasoning step, an executable tool invocation, and a natural language acknowledgment. Every response is designed to be parsed, validated, and executed by the DarajPay MCP server.

The model runs comfortably on a single Tesla T4 GPU and trains in minutes, which means we can iterate on it quickly and deploy it close to the users and systems it serves.

Why a dedicated tool-calling model

Most assistant workloads in a payments and ticketing platform are not open-ended. They are structured: register a user, search for buses from Jijiga to Addis, book a seat, confirm a payment. These are deterministic operations with well-defined inputs and outputs, and they do not require a trillion-parameter model to get right.

By narrowing the scope, we gain three things. First, reliability: a small model trained on a tight schema is easier to validate and constrain. Second, cost: a 270M model is dramatically cheaper to run at scale than a frontier model, which matters when every user interaction is a tool call. Third, locality: the model is small enough to run on modest hardware, keeping latency low and data close to home.

We are not trying to build a model that knows everything. We are building a model that does a specific, important job extremely well.

How it works

The model is trained on conversational traces that follow the DarajPay tool-call schema. Each assistant response contains a thought describing the user's intent, a tool call with a name and arguments, and a short text response acknowledging the action.

A system prompt anchors every interaction: the model knows it is the DarajPay assistant, that prices are integers in ETB cents, and that it must never initiate or authorize a payment without explicit user confirmation. These constraints are reinforced in training and enforced again at the application layer.

Because the output is structured JSON, the platform can validate every tool call against a schema before execution. Malformed outputs are caught, repaired, or rejected, and the model never has unchecked authority over a financial operation.

Training and cost

Daraj-270M was fine-tuned using parameter-efficient LoRA adapters, training only about one percent of the model's parameters. A short training run completes in roughly three to four minutes on a single T4, peaking at under eleven gigabytes of reserved memory.

This efficiency is the point. It means we can retrain the model when the DarajPay tool schema changes, when we add new offerings or regions, or when we discover a failure mode in production. Iteration is measured in minutes, not weeks.

The model is released under the Apache-2.0 license, with the base weights derived from Gemma 3. We believe small, focused models should be open and reproducible, so we have published a full training and deployment guide alongside the model.

Safety and guardrails

A model that sits in front of a payment platform carries real responsibility. Daraj-270M is designed to run behind multiple layers of protection: a JSON schema validator for tool arguments, a tool registry that rejects unknown tool names, application-level authorization for payment and booking actions, and retry logic for malformed outputs.

The system prompt forbids initiating or authorizing payments without explicit user confirmation, and the platform enforces the same rule independently. The model is never the sole authority for a money-moving action.

We are honest about the limitations of a 270M model. It can occasionally emit malformed JSON or invent arguments, which is why validation is mandatory and not optional. The model is a component in a system, not a replacement for one.

What comes next

Daraj-270M is our first tool-calling model, and it is a foundation rather than a finish line. We are expanding the tool schema, adding support for more regions and languages, and improving the model's reliability on ambiguous and out-of-domain requests.

We will continue to publish our models, our training recipes, and the infrastructure that connects them to real systems. We believe the future of useful AI is small, local, and deeply integrated into the workflows where people actually work.

If you are building on the DarajPay platform, or if you are interested in compact tool-calling models for your own domain, we would love to hear from you.

Questions about this work? [email protected]

More from Daraj

  • Introducing Daraj.

    We are an AI research and infrastructure company building small language models and the systems that put them to work.

  • Introducing DarajPay.

    An Ethiopian payment and ticketing platform built to make booking, paying, and traveling simpler for everyone.