Quick conclusion
I let a language model running on my own PC write the first drafts of blog articles. After two articles, my answer is: it is good enough for a first draft and the research before it, as long as a second pair of eyes checks every fact afterward.
This is still a small sample: two articles, and only one has a formal evaluation. Treat it as an initial observation, not a benchmark.
What this series covers
This is part 1 of 5. This article is about why I tried a local model and what I asked it to do. Part 2 covers how I measure these models. Part 3 covers a file-handling mistake. Part 4 covers my setup.
Why bother with a local model
Three reasons, in the order they matter to me.
Cost. I run several blogs. Drafting is the most repetitive part, and every draft spends paid usage.
Usage limits. Cloud plans have caps. When I hit one, the work stops, or I wait. A model on my own machine has no cap other than time.
Curiosity. I wanted to see how far a model I can run at home actually gets on real work, not on a leaderboard.
Offline use and privacy matter a lot when you run a model locally. They were not what led me to try one for drafting; cost, usage limits, and curiosity did.
What I hand over, and what I don’t
For blog work, I hand over two things: the research before a draft, and the first draft itself. The model is a Qwen model, and it works inside a coding-agent tool (Qwen Code), so it reads my notes, searches them, and writes a file.
I also use it with a second agent tool to tidy files outside the blog. File management is a separate use case; Part 3 will cover a file-handling mistake.
I do not hand over fact-checking or the decision to publish. The reason is in the next section.
What came back
The first test was a short article for a Japanese cocktail blog I run. A different agent then re-checked the draft against primary sources, and I read the result.
What went well:
- Length, dates, and steps matched the source article.
- Its search through the notes was accurate. It cited the right lines.
- When it fetched the wrong thing, it noticed, and fetched again.
- When it did not know which brand a recipe used, it asked instead of guessing.
What went wrong:
- It added a sensation that was not in my notes. It described how the sweetness lingered in the glass. My records said “rich” and nothing more. Plausible, and invented. This was the most serious problem.
- It made an arithmetic error. Its research notes had the right ratio. The article body got it wrong.
- It slipped Simplified Chinese characters into a note meant for another tool. Rare, but easy to miss, and a simple script caught it.
- It stated a quality claim with no source, and got the backstory of one ingredient wrong.
I made 7 edits to roughly 1,050 characters of body text. Most were wording: two awkward negations, one pretentious phrase, two word choices, and two brand names I wanted added. I did not need to change the tone or the overall approach.
The verdict from that review: usable for the research and first draft of a simple recipe article, with an independent fact-check behind it.
A second article, a different setting
Because the two articles used different model settings, this pair is not a controlled comparison of settings. It has no formal evaluation. The fact-check changed one spot: a line implying a backstory that was not in my records. A recount of every number found no difference.
The failure modes were not identical across the two.
How I use it now
The flow is short:
- The local model does the research and writes the first draft.
- A different agent re-checks every claim against primary sources.
- I read it and fix what is left.
One habit I want to add: have the model list every statement that is not in my records, next to the draft. In the first article it reported one invented detail on its own, so the habit seems possible. But a list the model writes about itself is a starting point, not a check. The second reader still reads everything.
How far two articles go
- The sample is two articles, and only one has a formal evaluation.
- Both were short and simple. A harder article might go differently.
- I think it falls short of the frontier models I use. I have not tested both on the same article, so I cannot quantify the gap.
FAQ
Can a local model replace a cloud model for blogging? Not in my experience so far. It handles research and a first draft. I would not let it publish without a fact-check.
Do you need a powerful PC? It depends on the model. Part 4 will cover my setup.
Is the draft ready to publish as written? No. It needs a fact-check and an edit, and it sometimes adds things that are not true.
Next steps
Part 2 is about how I measure these models: a 14-question benchmark, what it can tell you, and what it cannot. If you want the short version of this article: a local model gave me usable first drafts, and I still check everything.