🔢 Model Parameter Counter
Estimate trainable parameters for dense layers, convolutions, embeddings, and transformer blocks. Calculate model size from parameter count and precision.
Calculate Model Parameter Counter
What this model parameter counter calculates
Estimate the number of trainable parameters contributed by common neural-network components. The calculator supports dense layers, 2D convolutions, embeddings, and a transformer-style block approximation.
Parameter count estimates storage and part of training memory, but it does not directly predict quality, latency, or total training cost.
Model Parameter Counter formula
Assumptions and limitations
The result is an engineering estimate based on the values entered. AI models, tokenizers, runtimes, accelerators, cloud services, and provider billing rules differ. Validate important decisions with measured data from the exact model, hardware, framework, and pricing plan you intend to use.
Worked example
A dense layer from 768 to 3,072 units has 768 × 3,072 + 3,072 = 2,362,368 parameters.
How to use the result
Parameter count estimates storage and part of training memory, but it does not directly predict quality, latency, or total training cost.
- Start with representative production assumptions rather than best-case demos.
- Run a low, expected, and high scenario to understand the range.
- Record model version, pricing date, hardware, precision, and workload details.
- Replace assumptions with observed p50 and p95 measurements after testing.
Common mistakes to avoid
- Counting activations or optimizer states as trainable parameters.
- Using only parameter count to predict runtime without considering sequence length and hardware.
- Forgetting tied embeddings, grouped convolutions, mixture-of-experts routing, or shared weights.
Frequently asked questions
Do non-trainable buffers count as parameters?
They may count in model storage but are separate from trainable parameters. This calculator focuses on trainable weights and optional biases.
How many bytes does one parameter use?
FP32 uses 4 bytes, FP16/BF16 uses 2, INT8 uses 1, and packed INT4 is approximately 0.5 byte before metadata.
Is the transformer estimate exact?
No. It is a planning approximation; exact architectures differ in attention projections, gated MLPs, normalization, embeddings, and expert layers.
Related AI calculators
Methodology and privacy
This educational tool uses the formula and assumptions displayed on the page. Calculations run locally in your browser, and the page does not transmit the values you enter. Results are estimates rather than provider quotes, benchmark guarantees, financial advice, or capacity guarantees. Last methodology review: August 2, 2026.
Formula Explorer connections
Interpretation: parameter count measures stored learned values, not runtime operations. It becomes memory only after multiplying by precision bytes. Assumption: architecture-specific extras such as normalization, tied embeddings, experts and buffers must be included separately when present.