What can a 16 GB AMD Radeon actually run locally?
Copyable ROCm/HIP and Vulkan presets for 16GB Radeon cards, with measured evidence and raw benchmark receipts behind them.
$
preset catalogue
Start with a recipe we would actually run. Each card carries its hardware lane, backend (ROCm/HIP vs Vulkan), power profile, exact configuration, validated context, and compact evidence. Recommended and alternative presets carry reviewed evidence; candidates are reproducible configurations still awaiting that evidence, so their metrics remain n/a. The explorer below remains for comparisons, experiments, and archive provenance.
Loading preset catalogue…
Visible profiles
0
0 raw runs loaded
Raw runs
0
in current filters
Highest context
n/a
configured tokens
Best generation
n/a
generated tok/s
MTP setups
0
visible model cards
Archived imports
0
hidden by default
Default view groups rows by model/setup, then shows prompt results inside each model card. Enable raw runs to inspect repeated measurements.
Generation tok/s is output-token speed after the prompt is processed.
Prompt eval tok/s is input/prompt processing speed before generation starts. MTP/speculation is shown separately because it changes generation behavior.
$
benchmark results
$
lane summaries
engine lanes
| Engine | Rows | Best Decode | Best Context |
|---|
model families
| Family | Rows | Best Decode | Best Context |
|---|