Curiosity · AI Model

Code Llama 70B

Code Llama 70B is Meta's January 2024 code-specialized open-weights model — a continued pretrain of Llama 2 70B on ~1T tokens of code that set the open-source coding bar for most of 2024. It has since been passed by DeepSeek Coder V2, Qwen 2.5 Coder, and Codestral, but remains a historical reference and is still deployed in many pipelines.

Model specs

Vendor
Meta
Family
Code Llama
Released
2024-01
Context window
16,384 tokens
Modalities
text, code
Input price
$0.9/M tok
Output price
$0.9/M tok
Pricing as of
2026-04-20

Strengths

  • Open weights under the Llama 2 community license — commercial use allowed
  • Specialized Python variant for single-language workflows
  • Mature inference stack support — llama.cpp, vLLM, TGI
  • Infilling capability for IDE-style middle-of-line completion

Limitations

  • Superseded on coding benchmarks by DeepSeek Coder V2, Qwen 2.5 Coder, Codestral
  • Only 16K context — short vs. modern coders at 128K+
  • No tool-use fine-tuning — pure code completion / generation
  • Built on Llama 2 base — weaker general reasoning than Llama 3+ coders

Use cases

  • Legacy code completion services still running on Llama 2 era stacks
  • Python-specific assistants (via Code Llama 70B Python variant)
  • Educational fine-tuning and research baselines
  • Self-hosted IDE extensions in environments that cannot use newer models

Benchmarks

BenchmarkScoreAs of
HumanEval≈67%2024-01
MBPP≈62%2024-01

Frequently asked questions

Is Code Llama 70B still worth using?

Only for legacy pipelines or compatibility. For new work, DeepSeek Coder V2, Qwen 2.5 Coder 32B, or Codestral beat it on benchmarks and offer longer context windows.

What are the Code Llama 70B variants?

Three: the base model (for code generation and infilling), Code Llama 70B Python (specialized on Python data), and Code Llama 70B Instruct (chat-tuned for conversational coding).

Will Meta release a Llama 4 Code?

As of April 2026, Meta has not released a code-specialized Llama 4. Community fine-tunes and general Llama 4 models are being used for coding while a dedicated Code Llama 4 is expected.

Sources

  1. Meta — Code Llama 70B announcement — accessed 2026-04-20
  2. Hugging Face — codellama/CodeLlama-70b-Instruct-hf — accessed 2026-04-20