CodeLlama Instruct

weights

open(downloadable on HF)

data

closed

code

partial(inference utils, no training pipeline)

license

Llama-2-Community(custom Meta license, non-OSI

Code Llama's weights are downloadable under Meta's Llama 2 Community License, which is not OSI-approved. The license bounds rather than forbids commercial use, since anyone within its limits may use the model commercially, unlike a license such as CC-BY-NC or Mistral's Non-Production License that forbids commerce outright.

• https://huggingface.co/codellama/CodeLlama-13b-Instruct-hf recorded 2026-09-29

  Model card for 13B-Instruct. Repo tags include "license:llama2" and "gated":false, so the safetensors download without a barrier. The Training Data section says "trained and fine-tuned using the same data as Llama 2 with different weights", so no corpus is released; the card gives only transformers usage snippets. Searched the body for a training pipeline, corpus release and fine-tuning scripts: none.

• https://huggingface.co/codellama/CodeLlama-13b-Instruct-hf/raw/main/LICENSE recorded 2026-09-29

  The LICENSE file in the model repo is the "LLAMA 2 COMMUNITY LICENSE AGREEMENT" with a "non-exclusive, worldwide, non-" transferable royalty-free grant, an Acceptable Use Policy, a bar on using outputs to improve other LLMs, and a "greater than 700 million monthly active users" condition; custom Meta license, not OSI-approved.

• https://raw.githubusercontent.com/meta-llama/codellama/main/README.md recorded 2026-09-29

  README of meta-llama/codellama says "This release includes model weights and starting code for pretrained and fine-tuned Llama language models" and "This repository is intended as a minimal example to load" Code Llama models "and run inference"; the run instructions cover only example_completion.py, example_infilling.py and example_instructions.py via torchrun. Searched the README for a training pipeline, pretraining or fine-tuning scripts and data preparation code: none, so the code is inference-only starting code.

Adoption is measured as Hugging Face downloads summed across the four Code Llama Instruct checkpoints this tier covers, most of them for the 7B model. All four repositories are declared on the product, and downloads are combined across the sizes rather than read from one checkpoint alone.

• https://huggingface.co/api/models/codellama/CodeLlama-13b-Instruct-hf recorded 2026-10-07

  2,475 downloads in the trailing 30 days for codellama/CodeLlama-13b-Instruct-hf

• https://huggingface.co/api/models/codellama/CodeLlama-34b-Instruct-hf recorded 2026-10-07

  23,834 downloads in the trailing 30 days for codellama/CodeLlama-34b-Instruct-hf

• https://huggingface.co/api/models/codellama/CodeLlama-70b-Instruct-hf recorded 2026-10-07

  593 downloads in the trailing 30 days for codellama/CodeLlama-70b-Instruct-hf

• https://huggingface.co/api/models/codellama/CodeLlama-7b-Instruct-hf recorded 2026-10-07

  33,954 downloads in the trailing 30 days for codellama/CodeLlama-7b-Instruct-hf

Code Llama Instruct is a 2023-era code model that sits well below current coding models. Its Instruct sizes are not monotonic: the paper gives 13B 42.7% HumanEval / 49.4% MBPP, with 70B higher at 67.8% and 34B lower at 41.5%.

• https://arxiv.org/abs/2308.12950 recorded 2026-08-13

  abstract - "Code Llama reaches state-of-the-art performance among open models on several code benchmarks, with scores of up to 67% and 65% on HumanEval and MBPP"; 7B/13B/34B/70B Instruct variants

• https://arxiv.org/pdf/2308.12950 recorded 2026-08-14

  Code Llama paper full text. Table 2, "Code Llama pass@ scores on HumanEval and MBPP", Instruct rows - 7B 34.8% / 44.4%, 13B 42.7% / 49.4%, 34B 41.5% / 57.0%, 70B 67.8% / 62.2% (HumanEval pass@1 / MBPP pass@1). The per-size figures are in the PDF text, not on the abstract page.

• https://huggingface.co/codellama/CodeLlama-13b-Instruct-hf recorded 2026-08-13

  Code Llama 13B-Instruct model card linking arXiv:2308.12950 "Code Llama - Open Foundation Models for Code"; no benchmark table on the card