Compare

Loading community benchmarks…

Community benchmarks

Compare

Put two chips, runtimes, or runtime versions side by side. Everything else is held fixed or called out, so a ratio is worth exactly what the matched runs say.

Add your benchmark

Runs compared: 23
Like-for-like pair: 1
Runtimes available: 2

Compare

Side A

Side B

Model

Chip

Quantisation A

Quantisation B

Backend

Prefill size

Decode size

Cooldown

14 runs vs 9 runs · 1 like-for-like pair

Held fixedModel: Qwen3 4B Instruct 2507Prefill: PP512Decode: TG128

The sides differ in conditioning, decode workload, harness schema. Ratios may reflect benchmark protocol rather than the runtime.

Runtime comparison

BaseRT vs llama.cpp

Ratios read side A relative to side B. Geometric mean over 1 like-for-like pair.

Decode: 0.83×-17.0%
Prefill: 1.32×+31.6%

Side A

BaseRT

14 runs

2 contributors

Latest 20 Sep 2026

Decode · TG128: 83.7tok/s median
best 153.7 · worst 54.8
Prefill · PP512: 2,543tok/s median
best 11,246 · worst 727

Models: Qwen3-4B-Instruct-2507, Qwen3-4B-Instruct-2507-Q4, Qwen3-4B-Instruct-2507-Q8
Chips: Apple M1 Max, Apple M3 Ultra, Apple M4 Max, Apple M4 Pro, Apple M5 Pro, NVIDIA GB10
Versions: BaseRT 0.2.4, BaseRT 0.2.6
Quantisations: Q4, Q8
Backends: CUDA, Metal
Cooldown: off
Conditioning: warmup_only
Decode workload: TG128, TG128 @ 1 ctx
Harness schema: basert-benchmark-harness/1

Side B

llama.cpp

9 runs

1 contributor

Latest 11 Sep 2026

Decode · TG128: 134.3tok/s median
best 185.1 · worst 60.8
Prefill · PP512: 4,793tok/s median
best 5,581 · worst 2,566

Models: Qwen3-4B-Instruct-2507
Chips: AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XT (RADV NAVI31), Apple M5 Pro
Versions: llama.cpp b10809 (5266f24da)
Quantisations: Q4_K_M, Q8_0
Backends: BLAS + Metal, ROCm, Vulkan
Cooldown: off
Conditioning: runtime_native_warmup
Decode workload: TG128
Harness schema: computearena-measurements/1

What differs besides the runtime

Facet	BaseRT	llama.cpp
Model	Qwen3-4B-Instruct-2507, Qwen3-4B-Instruct-2507-Q4, Qwen3-4B-Instruct-2507-Q8	Qwen3-4B-Instruct-2507
Chip	Apple M1 Max, Apple M3 Ultra, Apple M4 Max, Apple M4 Pro, Apple M5 Pro, NVIDIA GB10	AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XT (RADV NAVI31), Apple M5 Pro
Backend	CUDA, Metal	BLAS + Metal, ROCm, Vulkan
Conditioning	warmup_only	runtime_native_warmup
Decode workload	TG128, TG128 @ 1 ctx	TG128
Harness schema	basert-benchmark-harness/1	computearena-measurements/1

Like-for-like pairs

Each row is a configuration present on both sides. Values are per-cell medians; ratios read BaseRT relative to llama.cpp.

Configuration	Decode A	Decode B	Ratio	Prefill A	Prefill B	Ratio	Runs A / B
Qwen3-4B-Instruct-2507 Apple M5 Pro	76.9	92.6	0.83×	3,379	2,568	1.32×	4 / 3

Side A · BaseRT

Top 14 of 14 runs by decode throughput.

Model / format	Chip / backend	Decode	Prefill	Contributor	Date	Report
Qwen3-4B-Instruct-2507-Q4 BaseRT 0.2.6Q4	Apple M3 Ultra Metal	153.7 TG128 @ 1 ctx	2,529 PP512	basecompute	20 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q4 BaseRT 0.2.6Q4	Apple M4 Max Metal	121.9 TG128 @ 1 ctx	1,711 PP512	basecompute	19 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q8 BaseRT 0.2.6Q8	Apple M3 Ultra Metal	118.3 TG128 @ 1 ctx	2,557 PP512	basecompute	20 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q8 BaseRT 0.2.6Q8	Apple M4 Max Metal	104.3 TG128 @ 1 ctx	1,723 PP512	basecompute	19 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q4 BaseRT 0.2.6Q4	Apple M1 Max Metal	91.3 TG128 @ 1 ctx	922 PP512	basecompute	20 Sep 2026	View Benchmark
basecompute/Qwen3-4B-Instruct-2507 BaseRT 0.2.4Q4	Apple M5 Pro Metal	90.4 TG128	3,331 PP512	arki05	11 Sep 2026	View Benchmark
basecompute/Qwen3-4B-Instruct-2507 BaseRT 0.2.4Q4	Apple M5 Pro Metal	89.4 TG128	3,359 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q4 BaseRT 0.2.6Q4	Apple M4 Pro Metal	77.9 TG128 @ 1 ctx	727 PP512	basecompute	20 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q4 BaseRT 0.2.6Q4	NVIDIA GB10 CUDA	73.7 TG128 @ 1 ctx	11,246 PP512	basecompute	19 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q8 BaseRT 0.2.6Q8	Apple M1 Max Metal	72.1 TG128 @ 1 ctx	925 PP512	basecompute	20 Sep 2026	View Benchmark
basecompute/Qwen3-4B-Instruct-2507 BaseRT 0.2.4Q8	Apple M5 Pro Metal	64.3 TG128	3,398 PP512	arki05	11 Sep 2026	View Benchmark
basecompute/Qwen3-4B-Instruct-2507 BaseRT 0.2.4Q8	Apple M5 Pro Metal	62.8 TG128	3,405 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q8 BaseRT 0.2.6Q8	Apple M4 Pro Metal	56.2 TG128 @ 1 ctx	731 PP512	basecompute	20 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507-Q8 BaseRT 0.2.6Q8	NVIDIA GB10 CUDA	54.8 TG128 @ 1 ctx	10,970 PP512	basecompute	19 Sep 2026	View Benchmark

Side B · llama.cpp

Top 9 of 9 runs by decode throughput.

Model / format	Chip / backend	Decode	Prefill	Contributor	Date	Report
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	AMD Radeon RX 7900 XT (RADV NAVI31) Vulkan	185.1 TG128	4,783 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	AMD Radeon RX 7900 XT (RADV NAVI31) Vulkan	180.3 TG128	4,793 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	AMD Radeon RX 7900 XT ROCm	147.8 TG128	5,364 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	AMD Radeon RX 7900 XT ROCm	147.1 TG128	5,337 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q8_0	AMD Radeon RX 7900 XT (RADV NAVI31) Vulkan	134.3 TG128	4,824 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q8_0	AMD Radeon RX 7900 XT ROCm	110.8 TG128	5,581 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	Apple M5 Pro BLAS + Metal	94.4 TG128	2,568 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q4_K_M	Apple M5 Pro BLAS + Metal	92.6 TG128	2,566 PP512	arki05	11 Sep 2026	View Benchmark
Qwen3-4B-Instruct-2507 llama.cpp b10809 (5266f24da)Q8_0	Apple M5 Pro BLAS + Metal	60.8 TG128	2,686 PP512	arki05	11 Sep 2026	View Benchmark