What programming language are LLMs written in?

Most LLMs are built primarily in Python, with the heavy number-crunching handled underneath by highly optimized code in C++ and CUDA. Python is the language researchers and engineers actually write in, but it's not doing the intense math itself — it's orchestrating libraries that are.

Why is Python the standard for LLMs?

The reason comes down to a split between convenience and speed. Python is easy to read and quick to experiment in, which is why the AI field standardized on it — you can try an idea in a few lines instead of a few hundred. The trade-off is that Python is slow for raw computation.

So it's used as the top layer where humans work, while the demanding math is pushed down to faster code. Researchers get a language that's pleasant to think in, without paying Python's speed penalty where it would actually hurt.

What actually does the heavy lifting?

The real training and inference run through frameworks that sit underneath Python:

• PyTorch and, historically, TensorFlow — the frameworks engineers call from Python, built on fast, low-level C++.
• CUDA — Nvidia's platform for running code on GPUs, which handle the massive parallel matrix operations an LLM depends on.
• The GPU hardware itself — thousands of cores doing the same math at once, which is what makes training at this scale possible.

So when you run an LLM in Python, Python is really the conductor, handing the demanding work to compiled GPU code.

Is an LLM even a program?

It's worth noting that the model itself isn't really a program in the usual sense — it's a giant set of learned numbers, called weights, that the surrounding code loads and runs. The code is the engine; the model is the trained weights it drives. That's why the same model can run through different software: the Python and C++ around it can change, but the numbers that make it "know" things stay the same.

Language choice is about the engine, not the model.