llama2.c vs llama3 from scratch

llama2.c

• Developers who want to see how LLM inference works in plain C
• Students learning the Llama 2 architecture end to end
• People experimenting with tiny models trained on narrow datasets like TinyStories
• Anyone who wants a minimal inference engine with no external dependencies

llama3 from scratch

• Learners who understand basic Python and want to see how a transformer works internally
• Engineers who read model code and want a concrete, tensor-shape-level explanation
• Students studying attention and positional encoding

llama2.c

• Inference is a single short C file with no dependencies, which makes it easy to read
• The README includes a full train-plus-inference workflow in one repo
• Small models trained on narrow data can produce surprisingly coherent output, as shown in the examples

llama3 from scratch

• Clear, shape-by-shape explanations with diagrams that make the math easier to follow
• Uses the official Meta weights and config, so the numbers correspond to the real model
• Short, readable code that is easy to modify and experiment with

llama2.c

• People who need to run 7B or larger models, since the README says the code only does float32 inference and 13B+ does not currently work
• People who need fast production inference, since the README describes it as educational and float32-only
• People who want a polished, supported product, since the README says the project is young and moving quickly

llama3 from scratch

• People who want to run Llama 3 for chat or apps; this is an educational walkthrough, not a production inference server, and the README does not cover serving
• Readers without basic Python or linear algebra, since the explanations assume familiarity with tensors and matrix shapes
• Anyone looking for a complete implementation; the README says the BPE tokenizer is intentionally skipped and the truncated text covers only the first layer's early steps

llama2.c

• The README says only fp32 inference is supported, so large models need a lot of memory and run slowly
• 13B and larger models do not currently work according to the README
• Code Llama support is incomplete and the README calls it buggy

llama3 from scratch

• You must obtain the Meta weights separately, which requires accepting Meta's download terms
• The README is truncated and says the tokenizer is not implemented from scratch, so it is not a full standalone system
• Relies on torch and tiktoken, and the weights file is large, so it needs substantial local resources