Amid the recent proliferation of generative artificial intelligence tools, Meta has officially released “Code Llama,” a large language model (LLM) to generate code, support code completion, and to help debug human-written work.

The model, which is a code-specialised version of Meta’s Llama 2 LLM having been trained on code-specific datasets, supports prevalent languages such as Python, C++, C#, PHP, Java, Typescript, and Bash, among others. On top of creating and aiding with code, it can also generate natural language about code.

In the same vein as GitHub Copilot and Amazon’s CodeWhisperer, Code Llama “has the potential to make workflows faster and more efficient for developers and lower the barrier to entry for people who are learning to code,” according to Meta. “The goal is to make developer workflows more efficient so that they can focus on the most human-centric aspects of their job, rather than repetitive tasks.”

Code Llama itself comes in three different sizes: 7 billion parameters, 13 billion parameters, and 34 billion parameters. As is the case with LLM AI models, the number of parameters is a measure of the size and complexity of the model. Each of these models were trained with 500 billion tokens (i.e. units) of code, alongside code-related data.

While the 7B and 13B models are comparatively faster due to their sizes and more appropriate for lower-latency tasks, including real-time code completion, the 34B model “returns the best results and allows for better coding assistance,” according to the company.

Further, in addition to the different sizes of Code Llama, Meta has created “Code Llama – Python,” and “Code Llama – Instruct.” The Python variation has been “fine-tuned on 100B tokens of Python code,” with the company citing that Python is the most benchmarked language for code generation. The Code Llama – Instruct variation, meanwhile, “has been fine-tuned to generate helpful and safe answers in natural language,” and is recommended to be used alongside Code Llama.

The company released Code Llama for both research and commercial use under the same community licence as Llama 2, which DIGIT also covered the release of last month.

Recommended reading

• Meta Announces Commercial Version of its AI Model
• Developers Embrace AI Training, Expectations for Python Low
• Meta Pushes to Encrypt Messenger Despite Regulatory Pushback

Getting by With a Little Help from AI

In the release post for Code Llama, Meta noted that “Programmers are already using LLMs to assist in a variety of tasks.”

On this, and just two months ago following a recent survey from GitHub, DIGIT reported on the findings that 92% of developers are using artificial intelligence coding tools, which Meta’s new Code Llama will now join the ranks of.

Breaking that figure down, 67% said they use them both in work and outside of it, 25% said they just use them at work, and 6% said they only used such tools when outside of work.

The majority — 70% of respondents — noted that AI tools offer them an advantage in their work.

Further, 41% said that AI coding tools can help with the prevention of burnout, which is certainly not uncommon for developers. Haystack Analytics, for instance, found that 83% of software devs are suffering from burnout.