Models

Loading community benchmarks…

← All models

Community benchmarks

Gemma 4 26B A4B IT

Submitted benchmarks for Gemma 4 26B A4B IT. Browse individual runs and their measurement settings, or open the report JSON for full benchmark details.

Compare chips Compare runtimes Add your benchmark

Submitted runs: 12
Chips tested: 6
Contributors: 3

Model identity

Grouped as google/gemma-4-26B-A4B-it. The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution. · Nothing in this report ties it to a published model file. Reports from CLI 0.1.0 record only a model name and a file hash, so the run is grouped by the name it reported until it is reconciled.. Artifact verification confirms published model bytes, not that a community-submitted benchmark was honestly executed. Hover or tap a label for what it means.

This model group collects runs reported as basecompute/gemma-4-26B-A4B-it, gemma-4-26B-A4B-it-cuda-q4mix, gemma-4-26B-A4B-it-cuda-q8, gemma-4-26B-A4B-it-Q4, gemma-4-26B-A4B-it-Q8. Individual results keep the package name and format they were submitted with.

View on Hugging Face ↗

Published files benchmarked

basecompute/gemma-4-26B-A4B-it/gemma-4-26B-A4B-it-Q4.base (opens in a new tab)BaseRTQ4· 6 runs
basecompute/gemma-4-26B-A4B-it/gemma-4-26B-A4B-it-Q8.base (opens in a new tab)BaseRTQ8· 4 runs

Benchmark results

Peak throughput is the highest reported value and may come from different runs. Compare model, quantisation, backend, and token counts before drawing conclusions. Model verification identifies published artifact bytes; community benchmark execution is not remotely attested.

Individual runs for Gemma 4 26B A4B IT. Throughput is in tokens per second; dates use your browser timezone. Hover a date for the exact local time.
Model / runtime / format	Chip / backend	Decode	Prefill	Peak process memory	Contributor

gemma-4-26B-A4B-it-Q8 BaseRTQ8 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M4 Max Metal	73.8 tok/s TG128 @ 1 ctx	1,679 tok/s PP512	3,955.6 MiB	basecompute	20 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q4 BaseRTQ4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M4 Max Metal	81.5 tok/s TG128 @ 1 ctx	1,664 tok/s PP512	3,451.3 MiB	basecompute	20 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q4 BaseRTQ4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M3 Ultra Metal	81.5 tok/s TG128 @ 1 ctx	2,422 tok/s PP512	3,449.5 MiB	basecompute	20 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q8 BaseRTQ8 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M3 Ultra Metal	75.3 tok/s TG128 @ 1 ctx	2,258 tok/s PP512	3,951.9 MiB	basecompute	20 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q4 BaseRTQ4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M1 Max Metal	51.5 tok/s TG128 @ 1 ctx	871 tok/s PP512	3,451.0 MiB	basecompute	20 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q8 BaseRTQ8 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M1 Max Metal	45.5 tok/s TG128 @ 1 ctx	877 tok/s PP512	3,955.2 MiB	basecompute	19 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-cuda-q8 BaseRTQ4 Nothing in this report ties it to a published model file. Reports from CLI 0.1.0 record only a model name and a file hash, so the run is grouped by the name it reported until it is reconciled.	NVIDIA GB10 CUDA	40.4 tok/s TG128 @ 1 ctx	6,045 tok/s PP512	10,586.0 MiB	basecompute	19 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-cuda-q4mix BaseRTQ4 Nothing in this report ties it to a published model file. Reports from CLI 0.1.0 record only a model name and a file hash, so the run is grouped by the name it reported until it is reconciled.	NVIDIA GB10 CUDA	45.3 tok/s TG128 @ 1 ctx	6,190 tok/s PP512	7,293.5 MiB	basecompute	19 Sep 2026	View Benchmark
gemma-4-26B-A4B-it-Q4 BaseRTQ4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	NVIDIA GB10 CUDA	44.0 tok/s TG128 @ 1 ctx	6,179 tok/s PP512	7,385.5 MiB	basecompute	19 Sep 2026	View Benchmark
basecompute/gemma-4-26B-A4B-it BaseRTQ8· default-q8 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M5 Pro Metal	51.4 tok/s TG128	3,253 tok/s PP512	17,830.8 MiB	arki05	11 Sep 2026	View Benchmark
basecompute/gemma-4-26B-A4B-it BaseRTQ4· default-q4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M5 Pro Metal	51.7 tok/s TG128	2,047 tok/s PP512	3,491.6 MiB	arki05	11 Sep 2026	View Benchmark
basecompute/gemma-4-26B-A4B-it BaseRTQ4· default-q4 The report's model SHA-256 matches a file published on Hugging Face, so the exact model bytes are known. This does not attest benchmark execution.	Apple M5 Max Metal	92.0 tok/s TG128	4,055 tok/s PP512	3,526.5 MiB	lukas	10 Sep 2026	View Benchmark