gemma-4-31B-it on AMD GPUs

235 results · 10 sources · 260 measurements

Grouped by reported model identity. Exact weights, quantizations and software can differ. How results are reviewed

Documented results

24 selected tests


View ReportCustom evaluation	gemma-4-31B-itINT4 AutoRound AWQWeights ↗	MMLU57 subjects · first item per subject1 concurrent request	91.23 %Accuracy · 52 / 57 · Exact-choice parser	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS EndeavourOS (Arch Linux) Engine SGLang Backend ROCm Harness scripts/eval/eval_and_chart.py Prompt / output ? / ? tokens Context 8192 Concurrency 1 Date 2026-04-17 Full record →
View ReportLocalMaxxing	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	435.25 tok/sPrefill · unknown	12.65 tok/sGeneration · peak · unknown	3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Arch Linux Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 538 / 256 tokens Context 794 Concurrency Not reported Date 2026-06-27 Full record →
View Reportllama-bench	gemma-4-31B-itBF16Model ref. ↗	Token speed512 input tokens · 0 output tokens · 0 existing KV tokens	—	428.07 tok/sPrefill · mean · per request	—	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Fedora Linux 43 (Workstation Edition) Engine llama.cpp Backend ROCm / rocm-7_2_3 Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date 2026-05-16 Full record →
View Reportllama-bench	gemma-4-31B-itUD-Q8_K_XLModel ref. ↗	Token speed512 input tokens · 0 existing KV tokens	—	203.16 tok/sPrefill · mean · per request	—	2 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Fedora 42 — Linux 6.15.9-201.fc42.x86_64 (Sat Aug 2 11:37:34 UTC 2025) Engine llama.cpp Backend vulkan_amdvlk Harness llama-bench Prompt / output 512 / ? tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View ReportLocalMaxxing	gemma-4-31B-itQ4_K_MModel ref. ↗	Token speed	—	—	54.73 tok/sGeneration · unknown	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Linux Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 3663 / 2048 tokens Context 262144 Concurrency Not reported Date 2026-04-29 Full record →
View Reportllama-bench	gemma-4-31B-itQ5_K_MModel ref. ↗	Token speed	—	751.77 tok/sPrefill · mean · per request	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 26.04 LTS Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	920.55 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXLayer split · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	851.56 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXLayer split · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend Vulkan Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →
View Reportllama-bench	gemma-4-31B-itIQ4_NL_L_AMDWeights ↗	Token speed	—	1,381.37 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXMulti-GPU · split unknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Not reported Engine llama.cpp Backend ROCm 6.4.4 Harness llama-bench Prompt / output 1024 / 0 tokens Context Not reported Concurrency 1 Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ4_K_MModel ref. ↗	Token speed	—	849.32 tok/sPrefill · mean · per request	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Linux (inferred from documented test rigs) (inferred) Engine llama.cpp Backend ROCm 7.2.3 Harness llama-bench Prompt / output 1024 / ? tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View ReportLocalMaxxing	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	534.98 tok/sPrefill · unknown	14.49 tok/sGeneration · peak · unknown	3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Arch Linux Engine llama.cpp Backend rocm Harness LocalMaxxing community speed test; exact client where reported Prompt / output 538 / 256 tokens Context 794 Concurrency Not reported Date 2026-06-27 Full record →
View Reportllama-bench	gemma-4-31B-itBF16Model ref. ↗	Token speed0 input tokens · 128 output tokens · 0 existing KV tokens	—	—	3.55 tok/sGeneration · mean · per request	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Fedora Linux 43 (Workstation Edition) Engine llama.cpp Backend ROCm / rocm-7_2_3 Harness llama-bench Prompt / output 0 / 128 tokens Context Not reported Concurrency 1 Date 2026-05-16 Full record →
View Reportllama-bench	gemma-4-31B-itUD-Q8_K_XLModel ref. ↗	Token speed128 output tokens · 0 existing KV tokens	—	—	15.6 tok/sGeneration · mean · per request	2 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Fedora 42 — Linux 6.15.9-201.fc42.x86_64 (Sat Aug 2 11:37:34 UTC 2025) Engine llama.cpp Backend vulkan_amdvlk Harness llama-bench Prompt / output ? / 128 tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ5_K_MModel ref. ↗	Token speed	—	695.98 tok/sPrefill · mean · per request	—	2 × R9700Layer split · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 26.04 LTS Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	—	22.97 tok/sGeneration · mean · per request	2 × RX 7900 XTXLayer split · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 0 / 128 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	—	24.42 tok/sGeneration · mean · per request	2 × RX 7900 XTXLayer split · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend Vulkan Harness llama-bench Prompt / output 0 / 128 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →
View Reportllama-bench	gemma-4-31B-itIQ4_NL_L_AMDWeights ↗	Token speed	—	1,494.88 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXMulti-GPU · split unknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Not reported Engine llama.cpp Backend ROCm 6.4.4 Harness llama-bench Prompt / output 2048 / 0 tokens Context Not reported Concurrency 1 Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ4_K_MModel ref. ↗	Token speed	—	743.82 tok/sPrefill · mean · per request	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Linux (inferred from documented test rigs) (inferred) Engine llama.cpp Backend ROCm 7.2.3 Harness llama-bench Prompt / output 4096 / ? tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View ReportLocalMaxxing	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	458.91 tok/sPrefill · unknown	13.99 tok/sGeneration · peak · unknown	3 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Arch Linux Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 538 / 256 tokens Context 794 Concurrency Not reported Date 2026-08-04 Full record →
View Reportllama-bench	gemma-4-31B-itBF16Model ref. ↗	Token speed2,048 input tokens · 0 output tokens · 32,768 existing KV tokens	—	167.2 tok/sPrefill · mean · per request	—	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Fedora Linux 43 (Workstation Edition) Engine llama.cpp Backend ROCm / rocm-7_2_3 Harness llama-bench Prompt / output 2048 / 0 tokens Context Not reported Concurrency 1 Date 2026-05-16 Full record →
View Reportllama-bench	gemma-4-31B-itUD-Q8_K_XLModel ref. ↗	Token speed2,048 input tokens · 16,384 existing KV tokens	—	106.21 tok/sPrefill · mean · per request	—	2 × R9700Multi-GPU · split unknown · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Fedora 42 — Linux 6.15.9-201.fc42.x86_64 (Sat Aug 2 11:37:34 UTC 2025) Engine llama.cpp Backend vulkan_amdvlk Harness llama-bench Prompt / output 2048 / ? tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ5_K_MModel ref. ↗	Token speed	—	703.4 tok/sPrefill · mean · per request	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 26.04 LTS Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 1024 / 0 tokens Context Not reported Concurrency 1 Date Not reported Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	960.35 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXLayer split · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →
View Reportllama-bench	gemma-4-31B-itQ8_0Model ref. ↗	Token speed	—	609.25 tok/sPrefill · mean · per request	—	2 × RX 7900 XTXTensor parallel · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS CachyOS x86_64 Engine llama.cpp Backend Vulkan Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency 1 Date 2026-08-23 Full record →

Quantizations and software

Quantizations: BF16, INT4 AutoRound AWQ, IQ4_NL_L_AMD, IQ4_NL_M_AMD, Q4_K_M, Q4_K_XL, Q5_K_M, Q8_0, Q8_K_XL, UD-Q4_K_XL, UD-Q8_K_XL
Engines: llama.cpp, SGLang

Explore all 235 results →

This is a selection across sources, not a ranking or a matched experiment. Reading benchmark reports

GPU benchmark pages: R9700 · RX 7900 XTX · Ryzen AI Max+ 395