[bmdpat]

Field journal

Blog archive

Every published note, with the newest work first. Use the topic links to stay inside one search intent.

  1. How to Calculate Local LLM 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.

    #local-llm#model-evaluation#ollama
  2. Why a Failed Local LLM Benchmark Row Still Matters

    A failed local LLM row marks the test boundary. My RTX 5090 report shows why quality, speed, and settings belong in one receipt.

    #local-llm#model-evaluation#ollama
  3. Mojo Is Open Source. What Local AI Builders Need to Know

    Modular put Mojo under Apache 2.0 with LLVM exceptions on August 18, 2026. Here is what the release changes for local AI builders, and what it does not.

    #local-llm#open-source#ai-infrastructure
  4. My local LLM eval hid four token caps

    My 5090-rig eval uses four output caps: 50, 180, 192, and 256 tokens. A score without the cap is not a model result. Record the cap on every row.

    #local-llm#ollama#hardware
  5. A GPU Driver Is Not a Local LLM Benchmark

    My 2026-08-20 RTX 5090 report lists driver 610.88 and 17 benchmark rows. That does not show the driver changed model speed.

    #local-llm#hardware#llm-benchmarks
  6. The 26B Model Hit the Cap. The 8B Finished.

    Same code task, same 512-token cap. Ollama's done_reason showed one local model finished and one got cut off. Tokens per second hid it.

    #local-llm#ollama#benchmarking
  7. A 27B Model Fit on an 8 GB GPU. It Was Slow.

    Qwen3.8-27B Q3_K_S loaded on an RTX 3070. VRAM used 7,435 of 8,192 MiB. Decode ran 2.07 tok/s. Fit on 8 GB is not a usable rate.

    #local-llm#ollama#hardware
  8. My Local LLM Writer Failed Its Own Word-Count Gate

    Gemma 4 26B wrote a 535-word draft against a 550-word floor. The gate blocked it, both rescue engines were down, and no bad post shipped.

    #local-llm#ollama#hardware
  9. Unload Local LLMs After Every Test

    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.

    #local-llm#ollama#5090-reports
  10. How I Benchmark Local LLMs Before I Trust Them

    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.

    #local-llm#benchmarking#hardware
  11. How I Test a 30B Local Model Before I Load It

    A 30B local model can fit on paper and still fail the job. This test plan checks memory, tool use, speed, and repeatability first.

    #local-llm#hardware#ollama
  12. When a 4B Local LLM Beats 26B on One Task

    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.

    #local-llm#model-selection#hardware
  13. Does Ollama Include That New llama.cpp Feature?

    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.

    #local-llm#ollama#llama-cpp
  14. My Local LLM Got Faster After It Passed the Tests

    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.

    #local-llm#hardware#ollama
  15. Prime Agent hit 95.5% on ARC-AGI-3. I did not install it.

    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.

    #ai-agents
  16. My local models refused zero of 50 security tasks

    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.

    #local-llm#ai-security#open-weights
  17. The faster local model run took 83x longer

    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.

    #local-llm#llm-benchmarks#hardware
  18. Chunk Size Is a Reliability Setting

    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.

    #local-llm
  19. Why Local LLM Benchmarks Need Power Data

    A local LLM benchmark that reports only tokens per second hides the operating cost. My RTX 5090 run shows what power data adds.

    #local-llm#ollama#5090-reports
  20. VRAM Fit Is Not Runtime Support

    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.

    #local-llm#5090-reports#ollama
  21. My 5090 benchmark was missing the field I needed most

    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.

    #local-llm#llm-benchmarks#hardware
  22. Search Old Results Before Publishing an LLM Test

    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.

    #local-llm#model-testing#ollama
  23. Build Local LLM Eval Data From Real Failures

    I show how I turn failed local coding runs into replayable eval rows with the prompt, model output, tests, route, and verifier result intact.

    #local-llm#model-evaluation#ai-agents
  24. How I Keep LLM Results Valid After a Driver Update

    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.

    #local-llm#llm-benchmarks#hardware