Field journal
Blog archive
Every published note, with the newest work first. Use the topic links to stay inside one search intent.
A local model test is not over when text appears. I unload the model, read idle VRAM, and record the result before I start another run.
A local LLM benchmark should end with a decision. I record task quality, tokens per second, VRAM, power, and failure state before I add a model.
A 30B local model can fit on paper and still fail the job. This test plan checks memory, tool use, speed, and repeatability first.
On one RTX 5090 workshop, a 4B model beat a 26B model on speed while both passed four code checks. Here is the model-selection rule I kept.
A llama.cpp release note does not prove that Ollama can use the feature. I trace the active runtime path, pin versions, and test the same workload.
A three-run RTX 5090 test showed why local LLM tuning must pair speed with fixed-task checks. One faster setting also repaired every test.
Prime Intellect's new agent beat the human baseline on ARC-AGI-3 at launch. I copied one idea, queued one benchmark, and installed nothing.
Hosted coding assistants have declined defensive security work. I ran 50 such tasks across five local models on my own hardware and counted zero refusals.
One llama3.1:8b run posted 188.16 tok/s and finished in 43.9 seconds. The next posted 153.12 and finished in 0.53. Tokens per second ranked them backwards.
Three native crashes destroyed seven hours of compute on a fully offline OCR run. The fix was not more parallelism. It was a smaller unit of work.
A local LLM benchmark that reports only tokens per second hides the operating cost. My RTX 5090 run shows what power data adds.
A local model can fit in VRAM, download cleanly, and still fail before the first token. My 5090 test adds runtime support as a separate gate.
A fresh Qwen3.5 9B run showed 84.94 tok/s, but the useful number was the 6,105 ms load phase. I added phase timings and capture time to the benchmark receipt.
An independent QA pass caught my second post about the same Ollama batch sweep. Here is the duplicate check I now run before publishing an LLM result.
I show how I turn failed local coding runs into replayable eval rows with the prompt, model output, tests, route, and verifier result intact.
A new GPU snapshot does not refresh an old benchmark. I show how I bind driver, runtime, workload, and timestamps to each local LLM result.
My RTX 5090 runs show why model loading, prompt ingestion, token generation, and task checks need separate measurements before production use.
My RTX 5090 measurements show how I budget VRAM for a local model, a real workload, and the GPU processes that must stay resident beside it.
One local LLM speed number hides the work behind it. My RTX 5090 sweep shows why short generation, long context, and code need separate rates.
LM Studio shipped Bionic, a full agent built on open models with local code projects, voice, and document work. The interesting part is not the app. It is what.
Hugging Face ran its breach forensics on an open-weight model on its own hardware because hosted APIs refused the requests. Here is the lesson for builders.
A local model pass rate can be true and useless at the same time. Here are the three ways local LLM benchmarks mislead you, drawn from real rows on my RTX 5090.
I make my local LLM show its source plan before it writes. The preview gate catches weak retrieval while the fix is still cheap and easy to inspect.
My smaller local model existed and ran, but it did not beat the baseline. Here is the promotion gate I use before changing a working local AI route.