Meta releases Code Llama

Meta released Code Llama, a family of code-specialised models built by further training Llama 2 on code-specific data. It shipped in four sizes — 7B, 13B, 34B and 70B parameters — with the two smaller variants also supporting fill-in-the-middle completion, and in three variants at each size: a base model, a Python-specialised version, and an instruction-tuned version for natural-language coding requests.

Meta released the weights under the same community licence as Llama 2, permitting both research and commercial use free of charge. On its own reported benchmarks, the 34-billion-parameter model scored 53.7% on HumanEval and 56.2% on MBPP — the company described this as the best result among open coding models at the time and roughly on par with ChatGPT.

The release extended Code Llama’s open-weight strategy from general-purpose language modelling into a specific, commercially valuable niche: AI coding assistance, then dominated by closed, subscription-priced tools such as GitHub Copilot. By publishing a free, locally runnable alternative competitive on a standard benchmark, Meta put price pressure on paid coding assistants even where its model did not clearly lead on capability, reinforcing the pattern in which Meta used open releases to commoditise capabilities that OpenAI, Google and Microsoft were monetising directly.