AMD-Pro-BenchCommunity

13,812 benchmarks in 32 GPU combinations documented

Qwen3-4B on AMD GPUs

20 results · 5 sources · 52 measurements

Explore all results

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

Documented results

20 selected tests
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed——
183 tok/sGeneration · per request
1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11 (10.0.26200)
Engine
llama.cpp
Backend
Vulkan
Harness
llama-bench
Prompt / output
0 / 16 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-06
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed—
4,113 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
Vulkan
Harness
llama-bench
Prompt / output
128 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
View
Reportllama-bench		Qwen3-4BQ8_0Model ref. ↗Token speed—
6,260.78 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows
Engine
LLAMA-DX / llama.cpp fork
Backend
Vulkan
Harness
llama-bench
Prompt / output
512 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-07-20
Full record →
View
ReportServing benchmark		Qwen3-4BBF16Weights ↗Token speed1 · halo-arena-v1
68.8 tok/sThroughput · aggregate
—
70.92 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
vLLM
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
8192
Concurrency
1
Date
2026-07-17
Full record →
View
ReportServing benchmark		Qwen3-4BQ4_K_MWeights ↗Token speed1 · halo-arena-v1
123.81 tok/sThroughput · aggregate
—
125.25 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
llama.cpp
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
32768
Concurrency
1
Date
2026-07-15
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed—
3,923 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11 (10.0.26200)
Engine
llama.cpp
Backend
Vulkan
Harness
llama-bench
Prompt / output
128 / ? tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-06
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed—
5,223 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
Vulkan
Harness
llama-bench
Prompt / output
512 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
View
Reportllama-bench		Qwen3-4BQ8_0Model ref. ↗Token speed——
122.6 tok/sGeneration · per request
1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows
Engine
LLAMA-DX / llama.cpp fork
Backend
Vulkan
Harness
llama-bench
Prompt / output
0 / 128 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-07-20
Full record →
View
ReportServing benchmark		Qwen3-4BBF16Weights ↗Token speed1 · halo-arena-v1
261.95 tok/sThroughput · aggregate
—
68.2 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
vLLM
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
8192
Concurrency
4
Date
2026-07-17
Full record →
View
ReportServing benchmark		Qwen3-4BQ4_K_MWeights ↗Token speed1 · halo-arena-v1
290.26 tok/sThroughput · aggregate
—
79.69 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
llama.cpp
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
32768
Concurrency
4
Date
2026-07-15
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed——
174.5 tok/sGeneration · per request
1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
Vulkan
Harness
llama-bench
Prompt / output
0 / 128 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
View
Reportllama-bench		Qwen3-4BQ8_0Model ref. ↗Token speed—
1,120.92 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows
Engine
LLAMA-DX / llama.cpp fork
Backend
DirectX 12
Harness
llama-bench
Prompt / output
512 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-07-20
Full record →
View
ReportServing benchmark		Qwen3-4BBF16Weights ↗Token speed1 · halo-arena-v1
849.69 tok/sThroughput · aggregate
—
58.71 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
vLLM
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
8192
Concurrency
16
Date
2026-07-17
Full record →
View
ReportServing benchmark		Qwen3-4BQ4_K_MWeights ↗Token speed1 · halo-arena-v1
614.29 tok/sThroughput · aggregate
—
50.45 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
llama.cpp
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
32768
Concurrency
16
Date
2026-07-15
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed—
4,072 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
ROCm / HIP
Harness
llama-bench
Prompt / output
128 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
View
Reportllama-bench		Qwen3-4BQ8_0Model ref. ↗Token speed——
71.92 tok/sGeneration · per request
1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows
Engine
LLAMA-DX / llama.cpp fork
Backend
DirectX 12
Harness
llama-bench
Prompt / output
0 / 128 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-07-20
Full record →
View
ReportServing benchmark		Qwen3-4BBF16Weights ↗Token speed1 · halo-arena-v1
1,638.69 tok/sThroughput · aggregate
—
57.45 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
vLLM
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
8192
Concurrency
32
Date
2026-07-17
Full record →
View
ReportServing benchmark		Qwen3-4BQ4_K_MWeights ↗Token speed1 · halo-arena-v1
565.1 tok/sThroughput · aggregate
—
20.29 tok/sGeneration · per request
1 × R9700Single GPU · 32 GB VRAM/GPU
Configuration
Memory
32 GB VRAM/GPU
OS
Not reported
Engine
llama.cpp
Backend
ROCm / HIP
Harness
radeonrun bench.py (source profile framework: llama-benchy)
Prompt / output
? / ? tokens
Context
32768
Concurrency
32
Date
2026-07-15
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed—
6,671 tok/sPrefill · per request
—1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
ROCm / HIP
Harness
llama-bench
Prompt / output
512 / 0 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
View
Reportllama-bench		Qwen3-4BQ4_K_MWeights ↗Token speed——
129.2 tok/sGeneration · per request
1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU
Configuration
Memory
16 GB VRAM/GPU
OS
Windows 11
Engine
llama.cpp
Backend
ROCm / HIP
Harness
llama-bench
Prompt / output
0 / 128 tokens
Context
Not reported
Concurrency
Not reported
Date
2026-08-07
Full record →
Quantizations and software

Quantizations: BF16, Q4_K_M, Q8_0
Engines: LLAMA-DX / llama.cpp fork, llama.cpp, vLLM

Explore all 20 results →

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

GPU benchmark pages: R9700 · RX 9070 XT