Grouped by reported model identity. Exact weights, quantizations and software can differ. How results are reviewed
Documented results
24 selected tests| View ReportGeneric LLM | gemma-4-12B-it | RadeonForge routerEasy /14 held-out queries | 100 %Score · n = 1 · Held-out routing policy; deterministic JSON/label checks | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
|---|---|---|---|---|---|---|
| View ReportCustom evaluation | gemma-4-12B-it | GSM8KFull1319questions;5shot8K configured context | 94.54 %Accuracy · n = 100 · Original GSM8K tagged answer extraction and strict string comparison | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | 1,184.95 tok/sPrefill · unknown | 36.66 tok/sGeneration · peak · unknown | 3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPUConfiguration
|
| View ReportGeneric LLM | gemma-4-12B-it | Token speed | — | ~500 tok/sPrefill · range lower · per request ~900 tok/sPrefill · range upper · per request | ~45 tok/sGeneration · per request | 1 × RX 7700 XTSingle GPU · 12 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | 955.8 tok/sPrefill · median · unknown | 110.7 tok/sGeneration · median · unknown | 1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | — | 31.3 tok/sGeneration · median · per request | 1 × Ryzen AI Max+ 395Single GPU · Unified memory · 96 GB RAMConfiguration
|
| View ReportGeneric LLM | gemma-4-12B-itConflicting evidence | RadeonForge routerEasy /14 held-out queries | 85.7 %Score · n = 1 · Held-out routing policy; deterministic JSON/label checks | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportCustom evaluation | gemma-4-12B-it | GSM8KFull1319questions;5shot8K configured context | 94.62 %Accuracy · n = 100 · Original GSM8K tagged answer extraction and strict string comparison | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | 944.83 tok/sPrefill · unknown | 20.19 tok/sGeneration · peak · unknown | 3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | — | 85.34 tok/sGeneration · median · unknown | 1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | — | 24.72 tok/sGeneration · unknown | 1 × Ryzen AI Max+ 395Single GPU · Unified memoryConfiguration
|
| View ReportGeneric LLM | gemma-4-12B-it | RadeonForge routerEasy /14 held-out queries | 71.4 %Score · n = 1 · Held-out routing policy; deterministic JSON/label checks | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportServing benchmark | gemma-4-12B-it | Token speedGuideLLM multi-turn conversation | 21.66 tok/sThroughput · mean · per request | — | — | 2 × R9700Tensor parallel · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | 231.32 tok/sPrefill · unknown | 18.41 tok/sGeneration · peak · unknown | 3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | 2,373.27 tok/sEstimated prefill rate · median · unknown | — | — | 1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | — | 24.75 tok/sGeneration · unknown | 1 × Ryzen AI Max+ 395Single GPU · Unified memoryConfiguration
|
| View ReportGeneric LLM | gemma-4-12B-it | RadeonForge routerPolicy-focused /21 held-out queries | 100 %Score · n = 1 · Held-out routing policy; deterministic JSON/label checks | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportServing benchmark | gemma-4-12B-it | Token speedGuideLLM multi-turn conversation | 19.71 tok/sThroughput · mean · per request | — | — | 2 × R9700Tensor parallel · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | — | 2,350.81 tok/sPrefill · unknown | 58.17 tok/sGeneration · unknown | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Latency505 input · 256 output | 212.79 msTime to first token · Unknown | — | — | 1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Token speed | 633.4 tok/sEstimated prefill rate · median · per request | — | — | 1 × Ryzen AI Max+ 395Single GPU · Unified memory · 96 GB RAMConfiguration
|
| View ReportGeneric LLM | gemma-4-12B-it | RadeonForge routerPolicy-focused /21 held-out queries | 76.2 %Score · n = 1 · Held-out routing policy; deterministic JSON/label checks | — | — | 1 × R9700Single GPU · 32 GB VRAM/GPUConfiguration
|
| View ReportServing benchmark | gemma-4-12B-it | Token speedGuideLLM multi-turn conversation | 16.42 tok/sThroughput · mean · per request | — | — | 2 × R9700Tensor parallel · 32 GB VRAM/GPUConfiguration
|
| View ReportLocalMaxxing | gemma-4-12B-it | Other LLM measurements505 input · 256 output | 237.86 input+output tokens/stotal token rate · Unknown | — | — | 1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPUConfiguration
|
Quantizations and software
Quantizations: BF16, FP8, NF4, Q4_K_M, Q8_0, QAT-Q4_0
Engines: custom, llama.cpp, Loculus Core, Transformers, Transformers / PEFT, vLLM
This is a selection across sources, not a ranking or a matched experiment. Reading benchmark reports
GPU benchmark pages: R9700 · RX 9070 XT · RX 7700 XT · Ryzen AI Max+ 395