FP16
Definition pending
We've cataloged "FP16" but haven't written a full definition yet. Definitions are hand-curated rather than auto-generated, so it takes time to cover every term.
Want this one prioritized? Email us and we'll bump it.
Practical example
FP16 (half precision) stores numbers in 16 bits — half the memory of FP32, nearly identical inference quality for modern models. A 7B model: FP32 = 28 GB, FP16 = 14 GB, Q4 = ~4 GB. FP16 is the default for GPU inference because it's the sweet spot: perfect quality, half the memory, full GPU acceleration.
Workflow example
Most GGUF models are quantized to Q4/Q5 for consumer GPUs, but if you have a 24 GB GPU, run FP16 (or F16 GGUF) for zero quality degradation. In Ollama, if you see "quantization": "F16" in the modelfile, you're running full precision. FP16 inference is 2× the VRAM of Q4 but preserves quality. For production where quality is critical (medical, legal), FP16 or BF16 is mandatory.