Local LLMs Compared: Polishing Modes for Voice Typing

One thing I love about Commander Flow: it doesn't lock me into a single configuration. The tray gives me one-click access to a mode for the current task — from light polishing on battery to the heaviest model for serious work over selected text. That's rare in consumer AI apps.

I'm a curious user, so over April I ran every mode against my real workload. Here's what I pick when, and why.

What modes are there

Six in the tray. Without RAM numbers and technical detail: just six points on a "how deep should processing go" slider. From "don't touch anything" (dictation only) to "run it through the largest model and get a dense result."

I tried each on my real tasks. Here's what fits where.

Test 1: polishing everyday dictation

I dictated dozens of fragments of "live speech" — with filler words, slips, mid-thought changes. Subjective evaluation: how naturally did the mode clean fillers, place punctuation, and preserve my intonation.

• Lite — handles everything I say quickly and habitually. Occasionally leaves an extra comma.
• Balanced — a noticeable step up in "humanness" of polishing. A great middle ground for laptops without a GPU.
• Smart — the best balance. Text reads as if I'd carefully written it.
• Enhanced polishing — not the deepest mode but the most responsive. A good pick when you're firing off many short messages.

The takeaway here is simple: for daily dictation, Smart wins if the hardware can take it. On weaker laptops, Balanced and Lite still deliver quality results.

Test 2: voice command over selected text → translate and rewrite in style

The hardest scenario: I dictate in my native language, select my text, and ask "translate to German, business tone." The mode has to translate and rewrite in the target tone, preserving facts.

• Lite, Balanced — handle short phrases. Sometimes lose an idiom on long ones.
• Smart — consistently nails business tone in multiple languages. My pick for emails.
• Prompting mode — best on multi-clause voice commands like "translate to formal style, address to a woman, add a closing paragraph requesting confirmation." When the instructions stack up, I switch here.

Test 3: long selected text (~1700 words of documentation)

This stresses the model's ability to hold context. I selected a big chunk of docs and asked "summarize in five bullets, highlight the numbers."

• Lite — handles it but occasionally drops nuance from "far" paragraphs.
• Balanced — consistently doesn't miss anything.
• Smart — best at preserving hierarchy of importance.
• Prompting mode — on very large selections, it surfaces things the lighter modes ignore. Worth using when the summary is critical.

Test 4: voice commands over code and tech comments

I'm picky here. I want technical terms to not get rewritten: kubectl stays kubectl, not "Kubernetes," and so on.

All modes handle this well — Commander Flow uses a terminology dictionary from settings internally, and any model honors it. On Smart and Prompting mode, comment polishing comes out cleaner and tighter, especially when the comment mixes natural language with technical identifiers.

This is an important architectural detail: Commander Flow isn't just a wrapper around a local model, it's a tone and terminology control layer on top of one. So even the lightest mode produces results noticeably better than a "raw" local model on its own.

What I ended up choosing

Default: Smart. Best quality at reasonable RAM. Covers most tasks — dictation, emails, chats, basic voice commands over selections.

Switch to Prompting mode when I need a complex multi-step command on a selection, long summaries, or fine-grained style work. The largest model lives there.

Switch to Lite when I'm on the road on battery. Lower power use, baseline quality — but an email to Mom still polishes nicely.

Enhanced polishing is my pick when I'm sitting and rapidly dictating lots of short chat messages. Responsive and clean.

Off stays for the rare case when I want raw speech-to-text with no processing (a brain dump into a file, for example).

What this changes for the product

Being able to switch modes from the tray with no restart is a huge degree of freedom. I can adapt the tool to the current task in literally two clicks.

Commander Flow is clearly built on the assumption that local models will keep evolving, and the user shouldn't be stuck with whatever was new on purchase day. The team regularly pulls new architectures into the modes as they appear.

"Good AI software ages slower than the hardware. Commander Flow is one of those cases."

Things to be aware of

Switching modes requires loading. Heavy → light is almost instant. Light → heavy needs a few seconds to load from disk. It's a rare operation, but if you want to "quickly try another," allow the wait.

Disk size. All mode components together take a meaningful chunk of disk. You can purge unused ones from settings, but on first round of "let's see what's in here" the size surprises.

The most powerful mode wants decent hardware. On a really old laptop, Prompting mode runs slowly. If your hardware is weak, stick with Smart or below.

Who actually needs this

Most users will never go switch modes. They'll install Commander Flow, see Smart works great, and be happy.

But for those who want to dig — there's plenty here to dig into, and the digging pays off. I built my config, switch by context, and it feels like very mature product work.

Going back to work now in Smart. Writing the next post.