CodeLlama 34B Instruct

Code Llama 34B Instruct is the instruction-following 34-billion-parameter member of Meta's Code Llama family of code models. According to its Hugging Face model card, Code Llama was trained between January 2023 and July 2023 on an offline dataset, and the Instruct variant is meant to be safer for code assistant and generation applications. The family is described in the paper Code Llama: Open Foundation Models for Code. It is an older model, and Meta's newer Llama generations have their own catalog entries.

Key features

• Code completion and infilling: the card lists code completion, infilling, instruction following and chat among the family's capabilities.
• Three variants: base Code Llama for general code synthesis and understanding, Code Llama - Python specialized for Python, and Code Llama - Instruct for instruction following and safer deployment.
• Open weights: distributed on Hugging Face under the Llama 2 Community License, after sharing contact information with Meta.
• Sizes: the card says the family ranges from 7 billion to 34 billion parameters, and its table also lists 70B models.

Model specs

The card does not state a context length, and it carries no benchmark figures. It points readers to Sections 3 and 4 of the paper for evaluation results.

Limitations

Meta says testing was done in English only, that the model may in some instances produce inaccurate or objectionable responses, and that use in languages other than English is out of scope. Developers should run their own safety testing and tuning for their application.

Use cases

• Self-hosted code assistants and code generation experiments.
• Fine-tuning and research on code models.
• Baselines when comparing open code models.

Getting started

1. Accept the license and share your details on the model card.
2. Load the model with Hugging Face Transformers in bfloat16, or use a quantized build, since a 34B model needs substantial GPU memory.
3. Use the Instruct prompt format and run your own safety checks before deployment.

FAQ

What license does Code Llama 34B Instruct use?

The Llama 2 Community License, according to the model card.

When was Code Llama trained?

Between January 2023 and July 2023, according to the card.

What is the difference between Code Llama, Python and Instruct?

The base model targets general code synthesis, the Python variant is specialized for Python, and the Instruct variant follows instructions and is intended for safer deployment.

Is there a newer open code model?

Yes. Newer options in this catalog include Qwen2.5-Coder 32B Instruct and the general-purpose Llama 3.1 70B Instruct.

Alternatives

• Qwen2.5-Coder 32B Instruct: a newer open coding model.
• Llama 3.1 70B Instruct: Meta's later general-purpose model.
• DeepSeek Coder V2.5: another open coding option.

Conclusion

Code Llama 34B Instruct was an influential open code model, and it is still useful as a baseline or a research reference. For new projects, read the model card and compare it with newer models in the models category.