Parameter Count (Model Size)

The total number of trainable numbers, or weights, in a model, usually written with B for billion or M for million.

Parameters are the numbers a neural network learns through training, mainly the weights and biases of each layer, and the parameter count is simply how many of them there are. 7B means 7 billion parameters; 300M means 300 million. Parameter count broadly determines a model's capacity, and how much compute, memory, and time training and deployment cost. Counting weights alone, memory can be roughly estimated as parameter count times bytes per parameter: OpenVLA, despite being called “7B,” actually has about 7.5 billion parameters, taking roughly 15GB loaded in bfloat16 (2 bytes per parameter), plus more at runtime. Robot models vary enormously in size: RT-2's largest version has 55 billion parameters, π0 has about 3.3 billion, and Hugging Face's SmolVLA has only 450 million. More parameters isn't always better: OpenVLA, with roughly a seventh of RT-2-X's 55 billion parameters, actually scored 16.5 percentage points higher in overall success rate across 29 tasks.

ExampleFigure's Helix splits its two parts at very different scales: the 7-billion-parameter vision-language model handling understanding runs only 7 to 9 times per second, while the 80-million-parameter policy producing actions runs 200 times per second.

Also called

Model Scale, Model Size, B (billion parameters), M (million parameters)

Related

Neural Network · Scaling Law · Inference Latency · Post-Training Quantization · On-device Model · Foundation Model

Sources

OpenVLA: An Open-Source Vision-Language-Action Model (arXiv:2406.09246)
RT-2: Vision-Language-Action Models (project page)
SmolVLA: Efficient Vision-Language-Action Model trained on LeRobot Community Data (Hugging Face blog)

As of

2025-06