Guides
Find a useful starting point.
Working with your own documents? Trying an agent? These are the write-ups we come back to, with examples from our own runs. Pick the job you want to do.
Put it to work
Documents and workflows
Embedding benchmark
Build a search index for your own documents
We indexed 769 documents on an M3 Ultra. See the chunking, embedding models, timings, and why avoiding a tiny cloud bill wasn’t the main reason to run locally.
Read the embedding testWorking notes
Give an agent a bounded job
A six-stage workflow for turning a request into a task, checking the result, and deciding what happens next. Useful when a long chat is getting hard to manage.
Read the workflow notesChoose with evidence
Model and hardware comparisons
A benchmark, a field report, and a coding test answer different questions. Each write-up explains what was run and where the comparison stops.
Image benchmark
Compare image generation on Mac and Spark
The same Qwen Image 2.1 prompts across M3 Ultra, M5 Ultra and DGX Spark. Compare the base and Turbo outputs alongside timing and runtime differences.
See the image comparisonAgent field report
Try a real agent task locally
Local Qwen, Muse and DeepSeek tackled a Hermes plugin inventory. The report covers waits, response formats, and the limits of comparing these runs.
Read the Hermes field reportCoding comparison
Check whether a model finishes the job
Five models took on one coding task. Four cloud models passed; the local 27B model didn’t. Read the task and its constraints before applying the result to your own setup.
Read the coding comparisonFor the full collection, browse Articles. Questions or a result that doesn’t look right? Meet the author and send a correction.