How much of a model is actually in VRAM?
Ollama exposes the loaded model, its quantization and its reported VRAM size in one local request.

In this article
The question
Two installations carrying the same model nickname can load different artifacts and context lengths. A reproducible observation starts by asking the runtime what is loaded. This is a measurement protocol; no GPU results are claimed.
Collect a record
With the model loaded on the local Ollama server, run curl http://localhost:11434/api/ps. Save the response and timestamp. Record the model digest, quantization level, context length and size_vram. If the models list is empty, send a normal request to load your intended model and repeat the check.
Repeat under one change
Keep the prompt and model digest fixed while changing a single setting. Repeat the observation with the same other applications running. Report the raw byte count and your unit conversion. The runtime’s model VRAM field and whole-device memory use describe different quantities; label each separately.
Measure the same model under two workloads
Start with one prompt of a few sentences. Record the exact model variant, GPU memory before loading, running-model status after inference and peak GPU memory while the request runs. Next, use a longer prompt and verify the configured context size. Observe how peak memory and speed change. Keep other GPU processes constant and separate first-request loading from a warmed run. A mixed CPU/GPU placement can make the model usable while changing latency substantially.
Now make one reversible change, such as reducing the configured context to match the task. Run the long case again and check that the answer still uses the information it needs. If memory demand drops but a critical fact disappears from the answer, your usable context is too small for that workload.
Report a result someone can reproduce
Write down GPU model, runtime version, quant filename, context setting, concurrent requests and sample size. Include evidence of actual placement when possible. Share the prompt or a representative fixture. Treat your results as measurements of the tested configuration. Another machine can differ even with the same label on the model download.
Compare Qwen3.8 GGUF downloads, including GSQ-RCO and ternary builds
Primary sources
Ollama API referenceRead next Read the number behind the model name