Stability AI, creators of Stable Diffusion, announced its first open LLM generative AI product for coding called StableCode.
According to Stability AI, StableCode operates in three distinct modes — a base model for general use, an instruction model, and a long-context-window model that can support up to 16,000 tokens — StableCode is tailored to meet various coding needs. Whether you’re a beginner or a seasoned programmer, StableCode has something for you.
The foundation of StableCode lies in the BigCode project, an open-source initiative that provided the initial set of programming language data. Stability AI took this data, filtered, and fine-tuned it to create a model that’s not just efficient but also innovative.
The use of rotary position embedding (RoPE) sets StableCode apart from other models, allowing for a more balanced approach to code generation without weighing current tokens more than past ones.
StableCode is not limited to one or two programming languages. It supports development in Python, Go, Java, JavaScript, C, markdown, and C++. This wide array of supported languages makes it a versatile tool for developers across different platforms.
One of the standout features of StableCode is its long-context-window version, which allows users to look at a medium-sized code base that includes multiple files. This helps in understanding and generating new code that’s tailor-made to your needs.
In benchmarks, StableCode matched or exceeded similar scale models like GPT-3.5 on programming tests. The long-context-window variant allows viewing more code simultaneously to potentially generate better solutions.
Christian Laforte, head of research at Stability AI, envisions StableCode as a tool that empowers everyone to become a programmer. Just as Stable Diffusion turned everyone into an artist, StableCode aims to enable anyone with a good idea to write a program that solves their problem.
In addition to version 1.0 of the SD XL image model, Stability AI released in April its first open-source language model, StableLM, and FreeWilly, a language model based on Meta’s Llama v2 and improved with a synthetic dataset. These match or somewhat exceed the performance of the original model and GPT-3.5 (ChatGPT).
