Open source AI models: top downloads and a hardware check

Open source vs. open weights: an important distinction

People often say “open source models” to mean AI models you can download and run yourself. The phrase is handy but loose. A truly open source model publishes its weights, its code and ideally information about its training data, under a license that allows use, modification and redistribution, including commercial use. Many models are actually open weights: the parameters can be downloaded, but the license may restrict certain uses, set conditions for large companies or forbid using the outputs to train another model.

The “license” column in the table above shows the license declared on Hugging Face, the go-to platform for sharing these models. Before any business use, read each model’s full license: that document is what counts, not the “open source” label.

How to read the table

The table lists the most downloaded text generation models on the Hugging Face Hub, with their license, size in parameters, downloads and likes. Downloads reflect real-world use; likes reflect community interest. A heavily downloaded model is not necessarily the best one for your task, so cross-check with the scores in the AI ranking and the model pages.

What is a parameter?

A parameter is a number adjusted while the model is being trained. Together, these numbers encode what the model has learned. Model size is expressed in billions of parameters, hence the “B” in names like “8B.” The more parameters a model has, the more capable it generally is, but the more memory and computing power it needs.

Quantization, simply explained

Each parameter is stored at a certain precision. At 16 bits it takes 2 bytes; at 8 bits, 1 byte; at 4 bits, half a byte. Quantization means lowering that precision to make the model lighter, a bit like compressing a photo: the file gets much smaller, at the cost of a quality loss that is often subtle at 8 bits and more noticeable at 4 bits, depending on the model and the task.

The calculator’s formula

The “will it run on my computer?” calculator uses this formula:

memory needed (GB) ≈ billions of parameters × bytes per parameter × 1.2

The 1.2 factor adds headroom for what sits alongside the weights while the model runs, mainly the conversation context and the software’s own needs. Example: an 8-billion-parameter model at 4 bits needs 8 × 0.5 bytes × 1.2 ≈ 4.8 GB.

The result is compared with the RAM or VRAM you enter. Green: you have enough memory. Orange: it is tight, and the model may be slow or limited. Red: not enough memory, so try stronger quantization or a smaller model.

RAM or VRAM?

RAM is your computer’s system memory; VRAM is the memory on your graphics card (GPU). A model loaded into VRAM runs much faster, because GPUs excel at the parallel calculations models rely on. Without enough VRAM, a model can run on the CPU using RAM, or be split between the two, but generation slows down. On computers with unified memory, such as some recent Macs, the CPU and GPU share the same pool: enter the total memory, keeping some for the operating system.

Tools for running a model locally

• Ollama: a tool that downloads and runs a model with a single command.
• LM Studio: a desktop app with a graphical interface, handy for finding a model, choosing its quantization and chatting without touching the terminal.
• llama.cpp: the open source engine many tools are built on, for those who want to fine-tune performance.

Pros and cons

The main benefit is privacy: your text never leaves your machine, which matters for personal or work documents. You also avoid per-token billing and can work offline. The limits are real: models that fit on a personal computer often trail the largest online models, speed depends on your hardware, and updates, security and power consumption are your responsibility.

To compare with an online model, use the AI agent cost calculator. To build your own agent on a local model, see the Build learning path and the agent frameworks page.