#rtx5090

Compare a local model with an API on the same task. Record settings, failed attempts, review time and cost before you decide where the workload belongs.

My local writer hit a host RAM guard, then Sonnet produced a draft that failed QA. Here is how I separate availability, model identity, and useful output.

A 94x bigger prompt cost llama3.1:8b only 9.7% of its generation rate. VRAM rose 656 MiB, but num_ctx doubled in the same step. Budget memory first.

gemma4:26b read 53 prompt tokens at 14.5 tok/s and 4,470 at 6,178.9 on the same RTX 5090 run set. Prompt speed is not one number.

On September 4, 2026 my llama3.1:8b daily brief was ready with 0 model calls and 0 seconds. Preview-only file ranking is not a model run.

My RTX 5090 run spent 43.48 seconds on model load and 0.17 seconds on output. Split wall time before you judge local LLM speed.

My RTX 5090 had 32 GB of VRAM ready, but my local policy stopped Ollama at 7.02 GiB of free host RAM. Here is my preflight before model load.

I audited seven RTX 5090 benchmark files on 2026-08-27. Not one records the GPU driver version. Here are the four fields every local LLM row needs.

I audited seven RTX 5090 benchmark CSVs and found four header widths, three power schemas, and no schema version. Here is the row contract I use now.

My local AI test stopped before model load because the runtime could not build. This preflight keeps setup failures out of benchmark results.

Gemma 4 26B Q4_K_M averaged 49 W on a long RTX 5090 run and peaked at 338 W. Keep both watt numbers before you compute energy per token.

My RTX 5090 test shows how watts and output rate become joules per token, and why the faster of two matched settings can waste energy.