I built a programming language.
It is called Quokka, and it is designed as an AI/ML-focused alternative to Python.
The goal was not to create another general-purpose language. Instead, I wanted to explore what a programming language could look like if machine learning workflows were treated as a first-class use case from the beginning.
Why Quokka?
Python has become one of the most important languages in machine learning and artificial intelligence. Its ecosystem is difficult to replace, particularly because of the enormous number of libraries and tools available.
However, that also creates an interesting question:
What would a Python alternative designed specifically around modern AI/ML workflows look like?
Quokka is my attempt at exploring that question.
Self-hosting
One of the core goals of Quokka is self-hosting.
The compiler includes bootstrap verification through:
quokka check-self
This allows Quokka to verify aspects of its own compiler toolchain rather than treating the compiler as an entirely external component.
For me, this was an important part of the project because a language should eventually be able to become increasingly independent from the tools that initially created it.
Language design
Quokka includes several features intended to make programs easier to reason about.
Immutable-by-default variables
Variables are immutable by default, reducing accidental state changes.
Result and Option-style error handling
Instead of relying exclusively on exceptions, Quokka provides Result/Option-style approaches for representing successful, unsuccessful, or optional values.
Python and ML interoperability
Quokka can communicate with Python and machine-learning workflows through JSON-based IPC.
This is important because Quokka is not intended to exist in isolation from the Python ecosystem.
Python has an enormous ML ecosystem, and interoperability makes it possible to experiment with Quokka without abandoning the tools that already exist.
Machine learning
Quokka is designed with machine learning workflows in mind.
The project already supports workflows involving QLoRA fine-tuning, including experiments with a 1B-parameter model.
Quokka can also work alongside tools such as Unsloth, allowing it to interact with existing ML infrastructure rather than requiring an entirely new ecosystem.
The project also includes Joey, Quokka's native ML and fine-tuning layer.
Developer tooling
Quokka currently includes:
- A Windows installer
- VS Code syntax highlighting
- Self-hosting and bootstrap verification
- Python/ML interoperability
- QLoRA support
- MIT licensing
The project is still early, so the ecosystem is nowhere near the maturity of Python. That is intentional transparency: Quokka is an experiment and an evolving language, not a claim that Python can simply be replaced overnight.
Built in India
Quokka was built in India as an independent project.
One of the things I find most interesting about programming languages is that they allow a developer to experiment with the foundations of software itself.
Instead of only building an application with existing languages and frameworks, I wanted to build the language underneath the application.
That is what led to Quokka.
What's next?
There is still a lot to build.
The next stages include improving the compiler, expanding the language, strengthening the ML tooling, improving documentation, and making the developer experience easier for new users.
I also want to continue exploring how much of a modern AI/ML workflow can be handled natively while still maintaining interoperability with the existing Python ecosystem.
Quokka is an early project, but it is now publicly available.
Quokka: https://quokka.space
The source code and project development are available publicly as the language continues to evolve.
If you are interested in programming languages, compilers, machine learning infrastructure, or experimental developer tools, I would be very interested in hearing your thoughts and technical feedback.
Top comments (0)