§ 001 / MEASURED FIT RECEIPT
RTX 5090 32GB runs Gemma 4 26B-A4B.
One recorded result from owned hardware. The number below is tied to the quant, context, workload, and measurement date shown.
GPU
RTX 5090 32GB
Model tag
gemma4:26b
Generation
198.8 tok/s
Measured
2026-07-09
§ 002 / MEASURED ON REAL HARDWARE
Headline run
Workload
Short generation · 256 tokens
Prompt eval
15 (cold) tok/s
Peak VRAM
19.9 GB
Source date
2026-07-09
Run command
ollama run gemma4:26b§ 003 / ALL RECORDED RUNS
All recorded runs
Context changes the memory budget and speed. A result measured at 4K is not a measured claim at 8K.
| Context | Workload | Quant | Generation | Prompt | Peak VRAM |
|---|---|---|---|---|---|
| 4K context | Short generation · 256 tokens | Q4_K_M | 198.8 tok/s | 15 (cold) | 19.9 GB |
| 4K context | Agent code task · 512 tokens | Q4_K_M | 207.4 tok/s | 197 | 20.2 GB |
| 8K context | Long-context summarize | Q4_K_M | 180.2 tok/s | 6,179 | 20.2 GB |
§ 004 / CITE THIS RUN
Cite this measured run
BMD Pat measured RTX 5090 32GB running gemma4:26b (Q4_K_M, 4096 context tokens, Short generation · 256 tokens): 198.8 generation tok/s and 19.9 GB peak VRAM on 2026-07-09. Source: https://bmdpat.com/receipts/rtx-5090/gemma4-26b
Runtime version and model revision are not recorded in this catalog row. Keep that limit with the citation. Speed and memory measurements do not establish answer quality.
§ 005 / EVIDENCE AND NEXT STEP
Evidence and next step
This receipt is a measured data point, not a promise that every runtime or driver behaves identically. The source record is kept below so the claim can be audited and extended with the next run.
- Reports/5090/benchmarks/2026-07-09-5090-sweep-llama8b-gemma26b.csv via /5090-reports/latest.md