Code Llama
Meta's 2023 code models on Llama 2, frozen but still downloadable
Overview
Code Llama is Meta's family of code models built on Llama 2, released in August 2023 in base, Python and Instruct variants from 7B to 70B parameters. Meta has not updated it since early 2024 and its GitHub repository is archived. The weights remain downloadable from Hugging Face under the Llama 2 Community License.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Code Llama
- Self-hosted code completion from 7B, 13B, 34B or 70B checkpoints
- Fill-in-the-middle insertion of code between existing lines (7B and 13B base and Instruct only)
- Python-specialised generation from the Code Llama - Python checkpoints
- Natural-language coding requests answered by the Instruct checkpoints
- A fixed 2023 baseline for code-model research and benchmarking
- A licensed starting point for fine-tuning on in-house code, within the Llama 2 Community License terms
Code Llama was Meta's answer to closed coding models in 2023: base, Python and Instruct variants from 7B to 70B, inputs up to 100,000 tokens, and fill-in-the-middle on the 7B and 13B models. It still downloads, gated on Hugging Face under the Llama 2 Community License, which permits commercial use with conditions, including a separate licence above 700 million monthly users.
It is a frozen release: the GitHub repository is archived, the checkpoints have not changed since March 2024, and Meta's developer site now leads with its Muse models, including the Apache-licensed Muse Glimmer. Useful as a stable research baseline or a small self-hosted completion model; not where to start for a coding assistant in 2026.
What we found
Code Llama is the family of code models Meta released on 24 August 2023, built by further training Llama 2 on code. It came in three flavours, a base model, a Python-specialised model and an Instruct model for natural-language requests, first at 7B, 13B and 34B parameters and then, from 29 January 2024, at 70B. Meta's announcement says the models saw 500 billion tokens of code (one trillion for 70B) and handle inputs of up to 100,000 tokens, and that only the 7B and 13B base and Instruct models were trained for fill-in-the-middle, the mode editor plugins use to insert code between existing lines. Meta reported 53.7% on HumanEval for the 34B model, which it called the best result among open models at the time.
Nothing has been added since. The meta-llama/codellama inference repository on GitHub is archived and read-only, with its last push in August 2024, and the twelve checkpoints on Hugging Face were last modified in March 2024. They are still there, gated: you accept Meta's terms and the Llama 2 Community License before downloading. That licence allows commercial use but is not an open-source licence in the usual sense. It carries an acceptable-use policy, bars using the outputs to improve other large language models, and requires a separate licence from Meta for any company with more than 700 million monthly active users.
Meta's developer site has moved on. llama.com now redirects to dev.meta.ai, which leads with the Muse family: Muse Wonder and Muse Spark on the Meta Model API, a terminal coding agent called Muse Code, and Muse Glimmer, a 30B open-weight model under Apache 2.0. Code Llama survives under "Other models" in the Llama prompt-format docs.
Treat it as a fixed, well-documented 2023 baseline for research, or a small self-hosted completion model where its licence suits. For a coding assistant today, a current open-weight model is the better starting point.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Code Llama supports code infill at 100,000 token context — long enough to fit over 2,000 lines of Python in a single context window.
- The Python-specialised variant was trained with a much higher proportion of Python code than the base model, outperforming the base on Python-specific benchmarks.
- Code Llama 70B was the first openly available model to pass 50% on the HumanEval coding benchmark, a major milestone for open-source code models.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Code Llama's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Code Llama unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Code Llama directly →
Spotted something out of date? Suggest an update →
Alternatives to Code Llama
More in Code & Dev Tools
