read first
A computer for local coding AI: autocomplete and agents need different tests
Coding can mean finishing the line under your cursor, explaining an error or changing several files. A single speed table cannot predict satisfaction with all three. Start by separating the jobs you want to delegate.
Does a line suggestion need your largest model?
Autocomplete must arrive before you type past it. A setup that produces useful short suggestions promptly can be more comfortable than a large model that excels at explanations. Try it in normal files and note accepted suggestions and interruptions. Check the editor and model's completion support: good code in a chat window does not guarantee good insertions at the cursor. Establish that narrow workflow before choosing hardware around it.
Make an edit task checkable
Prepare a reproducing input, current output and expected output for a bug fix. Supply the relevant function and caller rather than every unrelated file. Review the diff and tests, not the confidence of the explanation. Three failed attempts from a fast model may take longer than one successful attempt from a slower model. Do not turn one success into a universal intelligence ranking; keep several recurring tasks from your project and reuse them under the same conditions.
Do not start by sending the whole repository
As an agent reads files, earlier answers and tool output accumulate alongside code. Later requests can therefore be heavier than the first. File selection and context boundaries matter. A large context allowance does not guarantee the model chooses the right files. Start with one feature and inspect what the agent reads. Buying memory to fix poor context selection can increase spending without removing the cause; context size and retrieval quality are not interchangeable.
Tool permissions are a separate decision
Local inference is not a reason to grant unrestricted terminal and file access. Start in a copy or separate branch and review changes before applying them. Keep secrets and real customer data out of initial tests. Connected search tools and APIs may still transmit data. Executing code or installing packages has greater consequences than reading generated text, so decide which actions require approval. A useful coding environment makes mistakes visible and reversible rather than granting the broadest permissions.
Compare through the end of one task
Record first response, revision count and time through passing tests as well as generation rate. Autocomplete and file editing do not have to use the same model. Separate roles also change memory needs and the number of models kept resident. Use the site to compare estimated hardware speed with one model, then validate your normal task inside the editor. Streaming text is only part of the job; the useful configuration also reduces the time you spend reviewing and repairing it.