Numeric precision (FP16, BF16, FP8)
Numeric precision is the number format used to store a model's weights and run its maths, such as FP32, FP16, BF16, FP8 or INT4. Fewer bits per number means less memory and faster inference, at some risk to accuracy.
Also called: FP8, BF16, FP16, FP4, INT8, INT4, FP32, bfloat16
By Divya Sharma · Checked Nov 15, 2026 · How we review
Every weight in a model is a number, and the format of that number decides how much memory the model takes. A 16-bit format uses two bytes per weight, an 8-bit format one byte, a 4-bit format half a byte. That is why precision is the first thing to check when you work out whether a model fits in your VRAM; the GPU calculator does this sum for you.
The common formats
FP32 is full single precision, used mostly in training. FP16 halves the size but has a narrow range, so very large or small values can overflow. BF16 (bfloat16, designed at Google) keeps the same 8 exponent bits as FP32 and drops mantissa bits instead, so it has FP32's range with less detail. That makes it the usual default for training and serving modern models.
FP8 goes down to one byte. The FP8 standard defines two variants: E4M3, with more precision, and E5M2, with more range. Recent data-centre GPUs run FP8 natively. FP4 formats push further on the newest hardware.
INT8 and INT4 are integer formats. Weights are mapped onto a small set of whole numbers plus a scale factor. This is what quantization usually means in practice, and it is what files such as GGUF store for local use.
Trade-offs
Lower precision saves memory and often raises tokens per second, but each step down loses some information. The effect depends on the model and the method: 8-bit is usually close to the original, while 4-bit and below can hurt reasoning or long-context tasks. Hardware support also matters; a format your GPU cannot compute natively may be converted on the fly and give little speed-up.
Related terms
- Quantization
- Quantization is a technique that stores a model's weights, and sometimes activations, with fewer bits, such as 8 or 4 instead of 16, so the model needs less memory and often runs faster, at some cost in accuracy.
- VRAM
- VRAM is the memory on a graphics card. For AI work it holds the model weights, the KV cache and working data, and whether a model fits in VRAM usually decides which GPU you need to run it.
- GGUF
- GGUF is a binary file format for storing a model's weights, tokenizer and metadata in one file, built for fast loading and local inference with GGML-based tools such as llama.cpp, often in quantized form.
- Model weights
- Model weights are the learned numbers inside a neural network, set during training, that decide how it turns input into output. A checkpoint is a saved file or set of files containing those weights.
- GPU
- A GPU (graphics processing unit) is a processor with many small cores built to run many calculations in parallel. Its speed at matrix maths makes it the standard hardware for training and running AI models.