A faster GPU is only part of the answer. VRAM can change which models, quantizations, and context sizes are practical. UOS keeps capability, measured speed, system fit, and purchase evidence separate so one benchmark never hides the real trade-off.
An independent test measured multiple model/runtime configurations on one NVIDIA GeForce RTX 5080. These are runtime measurements on a fixed GPU, not GPU-vs-GPU scores, local FPS, or a hardware upgrade recommendation.
Each capture and condition is shown separately. Host CPU and OS were not published and remain unknown. GPU memory is the source-reported runtime allocation in MiB, not the card’s capacity.
Capture 2026-08-20
Ollama 0.32.1 vs llama.cpp b10507 · RTX 5080 · same capture is never joined to another date
Methodology: decode median of five runs; long-prefill uses the reported median. Settings and sample details are shown per runtime row.
gpt-oss-20b · capture
The test identifies different model conversions in this group. These are separate runtime configurations, not an identical-artifact comparison.
Context 8,192 · temperature 0 · seed 42 · GPU offload not reported
Decode median
186.26 tokens/sec
Long-prefill median
5,241.3 tokens/sec
GPU memory during run
12,310 MiB
Decode samples
5 runs
Show raw result rowgpt-oss-20b,lms,LM Studio 0.4.21,[email protected],defaults,8192,186.26,184.34,5241.3,474,78.9,12310,1119,8.43,0.36,12917.9,5,183.42|186.26|186.84|186.98|185.4
gpt-oss:20b
ollama ollama 0.32.15
defaults
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload not reported
Decode median
213.54 tokens/sec
Long-prefill median
9,239.4 tokens/sec
GPU memory during run
13,570 MiB
Decode samples
5 runs
Show raw result rowgpt-oss:20b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,defaults,4096,213.54,212.03,9239.4,475,,13570,1053,18.56,23.67,2138.8,5,213.22|213.64|213.48|213.54|213.76
gpt-oss-20b · normalized
The test identifies different model conversions in this group. These are separate runtime configurations, not an identical-artifact comparison.
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
242.83 tokens/sec
Long-prefill median
9,617.7 tokens/sec
GPU memory during run
12,533 MiB
Decode samples
5 runs
Show raw result rowgpt-oss-20b,lms,LM Studio 0.4.21,[email protected],normalized,4096,242.83,240.13,9617.7,474,38.6,12533,1101,8.43,0.27,13176.6,5,242.83|242.82|243.14|241.74|244.35
gpt-oss:20b
ollama ollama 0.32.15
normalized
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
213.65 tokens/sec
Long-prefill median
9,187.1 tokens/sec
GPU memory during run
13,570 MiB
Decode samples
5 runs
Show raw result rowgpt-oss:20b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,normalized,4096,213.65,211.8,9187.1,474,,13570,1053,8.05,8.36,2152.1,5,213.62|213.65|213.68|213.25|213.73
llama-3.2-3b · defaults
The model artifact is source-matched, but test settings are not fully verified as matched. Read each row's context and offload details before comparing these configurations.
Context 8,192 · temperature 0 · seed 42 · GPU offload not reported
Decode median
314.86 tokens/sec
Long-prefill median
20,341.8 tokens/sec
GPU memory during run
4,260 MiB
Decode samples
5 runs
Show raw result rowllama-3.2-3b-q4km,lms,LM Studio 0.4.21,[email protected],defaults,8192,314.86,308.68,20341.8,440,9.4,4260,1122,2.75,0.07,3471.9,5,273.28|312.54|316.75|315.22|314.86
llama3.2:3b
ollama ollama 0.32.15
defaults
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload not reported
Decode median
351.22 tokens/sec
Long-prefill median
17,802.7 tokens/sec
GPU memory during run
3,760 MiB
Decode samples
5 runs
Show raw result rowllama3.2:3b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,defaults,4096,351.22,345.19,17802.7,434,,3760,1053,2.67,2.79,1234.0,5,351.15|350.9|351.86|351.32|351.22
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
316 tokens/sec
Long-prefill median
20,336.7 tokens/sec
GPU memory during run
3,808 MiB
Decode samples
5 runs
Show raw result rowllama-3.2-3b-q4km,lms,LM Studio 0.4.21,[email protected],normalized,4096,316.0,309.91,20336.7,442,9.2,3808,1103,2.76,0.07,3469.6,5,310.55|316.0|315.86|316.23|316.6
llama3.2:3b
ollama ollama 0.32.15
normalized
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
350.52 tokens/sec
Long-prefill median
17,862.3 tokens/sec
GPU memory during run
3,760 MiB
Decode samples
5 runs
Show raw result rowllama3.2:3b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,normalized,4096,350.52,338.11,17862.3,432,,3760,1057,2.27,2.38,1239.6,5,349.35|350.38|350.52|350.63|350.96
qwen2.5-14b · defaults
The model artifact is source-matched, but test settings are not fully verified as matched. Read each row's context and offload details before comparing these configurations.
Context 8,192 · temperature 0 · seed 42 · GPU offload not reported
Decode median
98.15 tokens/sec
Long-prefill median
5,195.8 tokens/sec
GPU memory during run
11,194 MiB
Decode samples
5 runs
Show raw result rowqwen2.5-14b-q4km,lms,LM Studio 0.4.21,[email protected],defaults,8192,98.15,97.31,5195.8,454,24.6,11194,1103,7.38,0.15,10406.8,5,97.36|98.14|98.17|98.15|98.17
qwen2.5:14b
ollama ollama 0.32.15
defaults
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload not reported
Decode median
102.25 tokens/sec
Long-prefill median
4,983.8 tokens/sec
GPU memory during run
10,356 MiB
Decode samples
5 runs
Show raw result rowqwen2.5:14b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,defaults,4096,102.25,101.06,4983.8,453,,10356,1053,5.51,5.67,1604.1,5,102.13|102.34|102.25|102.3|102.25
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
97.63 tokens/sec
Long-prefill median
5,566.5 tokens/sec
GPU memory during run
10,405 MiB
Decode samples
5 runs
Show raw result rowqwen2.5-14b-q4km,lms,LM Studio 0.4.21,[email protected],normalized,4096,97.63,96.62,5566.5,453,24.7,10405,1106,7.68,0.16,10411.1,5,96.89|97.75|97.46|97.79|97.63
qwen2.5:14b
ollama ollama 0.32.15
normalized
Engine / backend: vendored llama.cpp 9d77fa172
Context 4,096 · temperature 0 · seed 42 · GPU offload full
Decode median
102.21 tokens/sec
Long-prefill median
5,038.3 tokens/sec
GPU memory during run
10,356 MiB
Decode samples
5 runs
Show raw result rowqwen2.5:14b,ollama,ollama 0.32.15,vendored llama.cpp 9d77fa172,normalized,4096,102.21,101.23,5038.3,452,,10356,1053,5.52,5.68,1617.6,5,102.21|102.21|102.17|102.21|102.21
Coverage scope: these counts represent Local AI decision inputs available in this view, not the complete benchmark or retail catalog. A verified requirement profile is not a measured GPU execution result.
Start with your hardware
Your PC is the baseline.
Save the GPU, CPU, memory, and motherboard you already own. Then choose Local AI as the goal and evaluate a candidate against the system you actually have.
Model fitDoes the selected model/profile fit the reviewed VRAM requirement?Capability
Measured speedUse exact GPU, model, backend, and context measurements where they exist.Performance
System fitPower, connectors, clearance, and supporting parts stay separate from AI speed.Compatibility
Purchase callPerformance and capability are combined with price only when compliant market evidence exists.Decision
How to read Local AI evidence
Capability gain and performance gain are different.
A 16 GB → 32 GB GPU can expand practical model and context choices even when the measured speed gain is smaller than the capability gain. UOS should report both rather than collapse them into one score.
What remains unknown
If an exact GPU/model/backend/context combination has not been measured, UOS does not substitute a sibling GPU or extrapolate a tokens-per-second number. Missing evidence remains research backlog.