Code Like a Ninja: Meet StarCoder 2, the Open-Source AI That Writes Code for You (Almost)

Developers have recently been using AI-powered code generators like GitHub Copilot and Amazon CodeWhisperer, but these tools often come with limitations like cost or restrictive licenses. To address this, Hugging Face and ServiceNow created StarCoder, an open-source code generator initially released in 2023. Now, the sequel, StarCoder 2, has arrived.

Unlike its predecessor, StarCoder 2 isn't a single model, but a family of three: a 3-billion-parameter version from ServiceNow, a 7-billion-parameter model from Hugging Face, and a 15-billion-parameter powerhouse trained by Nvidia. These models are accessible, running on most modern consumer-grade GPUs.

Similar to other code generators, StarCoder 2 can complete code lines, summarize existing code, and answer natural language queries about code. Notably, it boasts 4x the training data compared to the original, translating to significant performance improvements according to its creators. Additionally, the model can be "fine-tuned" using a GPU for tasks like creating chatbots or personalized coding assistants.

However, code generation raises concerns beyond speed and efficiency. A Stanford study highlights potential security vulnerabilities in applications developed using such tools. The survey also reveals anxieties around a lack of understanding regarding code generation processes and excessive code generation leading to "code sprawl."

StarCoder 2's licensing attempts to eradicate the “code sprawl” with BigCode Open RAIL-M 1.0 license but it also presents a potential hurdle. The license isn't truly "open" in the traditional sense as it restricts certain uses and may conflict with emerging AI regulations. However, Hugging Face maintains that it was carefully crafted to comply with current laws.

So, how does StarCoder 2 compare to the competition? It reportedly performs better than Code Llama 33B on specific code completion tasks while achieving double the speed. Additionally, being open-source allows developers to deploy the models locally and customize them for their specific needs, addressing concerns about privacy and security risks associated with cloud-based AI.

Furthermore, StarCoder 2 prioritizes ethical considerations and transparency. Unlike some competitors that train on potentially copyrighted code, StarCoder 2 relies solely on data licensed from the Software Heritage, a non-profit organization dedicated to code archiving. Additionally, the training data is readily available for developers to inspect, replicate, or audit.

While not perfect, with potential biases and limitations in handling non-English languages and specific coding languages, StarCoder 2 represents a step forward in the field. Its emphasis on transparency and open-source nature makes it a contender in the evolving landscape of code generators, offering developers a choice while fostering trust and accountability within the AI space.

As Star Coder 2 takes center stage, we know one thing for sure, AI is transforming into a full-blown tech revolution and everything in the coming two decades will be heavily based on automation through AI. If you want to leverage this to boost your business’s potential, look no further. We at Deqode are pioneer in helping companies elevate to their best in no time.

Contact us today to unlock the power of cutting-edge technologies and gain a competitive edge in this exciting frontier of technology.

Subscribe to The Deqode Digest to get weekly updates about the latest tech trends around you.?

Follow us on X for regular tech updates.

For users:

Hugging Face, which offers model implementation consulting plans, is providing hosted versions of the StarCoder 2 models on its platform. So is Nvidia, which is making StarCoder 2 available through an API and web front-end.

For devs expressly interested in the no-cost offline experience, StarCoder 2 — the models, source code and more — can be downloaded from the project’s GitHub page.