Model benchmarks

gemma-4-26B-A4B-it-qat-UD-Q4_K_XL local LLM performance

As of September 2026, gemma-4-26B-A4B-it-qat-UD-Q4_K_XL runs at up to 76.8 tok/s for local inference (best of 2 community benchmark runs across 1 GPU).

llama.cppQ4_K
ShareRedditX

Model size

26B

Peak speed

76.8 tok/s

Average speed

70.8 tok/s

Avg PP

647.6 tok/s

Min memory

12.7 GB

Max context

65.536 tokens

Avg output / run

4.935 tokens

Avg runtime / run

1m 9s

Avg quality

67.0

Benchmark runs

2

GPUs tested

1

Quality by task

Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 2 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.

TaskP5 (low)AvgP95 (high)
Overall64.967.069.1
Agent Workflow57.466.475.5
Code Generation58.661.263.8
Role Play & Narrative79.079.880.6
Research & Analysis57.860.663.4

Performance by hardware and tool

Every hardware/tool/quantization combination gemma-4-26B-A4B-it-qat-UD-Q4_K_XL has been benchmarked on, ranked by peak token generation speed. Last updated September 2026.

HardwareToolQuantBest tok/sAvg tok/sMemoryContextQualityRuns
Radeon Instinct MI60llama.cppQ4_K76.8 tok/s70.8 tok/s12.7 GB65.536 tokens67.02

Benchmark runs

All 2 gemma-4-26B-A4B-it-qat-UD-Q4_K_XL runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Frequently asked questions

Is gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf good for coding?
In our benchmarks, gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf scores 61.2/100 for coding. It runs at about 70.8 tok/s, so if you want more speed, unsloth/Qwen3.8-27B-GGUF:IQ3_S is faster (~104.3 tok/s) and still scores well for coding (83.0/100).
Is gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf good for agentic (tool-using) tasks?
In our benchmarks, gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf scores 66.4/100 for agentic workflows. It runs at about 70.8 tok/s, so if you want more speed, Tiel-Coder-35B-A3B-MTP-UD-IQ3_XXS is faster (~102.1 tok/s) and still scores well for agentic workflows (88.7/100).
How fast is gemma-4-26B-A4B-it-qat-UD-Q4_K_XL for local inference?
Across 2 community benchmark runs, gemma-4-26B-A4B-it-qat-UD-Q4_K_XL reaches up to 76.8 tok/s and averages 70.8 tok/s, with the fastest results on Radeon Instinct MI60.
How much memory does gemma-4-26B-A4B-it-qat-UD-Q4_K_XL need?
The leanest observed configuration used about 12.7 GB of memory (quantizations tested: Q4_K).
Which tools have been used to run gemma-4-26B-A4B-it-qat-UD-Q4_K_XL?
Benchmarks were submitted using llama.cpp. Results are community-contributed and updated as new runs arrive.