Outlier Core 27B vs Claude Opus — local vs cloud, by task

Outlier Core 27B on a Mac, locally, versus Claude Opus in the cloud. Two different bets on the same job. I spent time running the same kinds of prompts through both — reasoning, knowledge, writing, translation, and code — to see where a small local model genuinely keeps up and where the big cloud model still earns its keep. Here's the honest read.

Where they feel the same

For the bulk of everyday work, the gap is small enough that you stop noticing it. Explain a concept. Draft a blurb to a tight word count. Translate a paragraph faithfully. Refactor a function into something cleaner. On tasks like these, Core 27B's answers are the kind you'd ship without a second look, and in a blind comparison you'd struggle to pick which model wrote which.

That holds across categories — general reasoning, factual knowledge, prose, translation, and routine coding all land in the same neighborhood. The everyday floor for a good local model is higher than most people expect.

Where Claude still wins

I'm not going to pretend the gap closes everywhere. Claude beats Core 27B in a few clear spots:

• Speed. Cloud decode is several times faster end to end. A long answer that lands in about a minute in the cloud takes noticeably longer on your Mac. You feel that in a tight iteration loop.
• Novel research questions and very long contexts. On genuinely hard, open-ended reasoning and on contexts past 50k tokens, the big cloud models still pull ahead.
• Vision quality. Claude's image stack is the stronger one today.

The hard cases

The interesting question is what happens on the difficult prompts — the ones where a small local model should fall apart if it's going to fall apart anywhere. Things like a chess engine with the full rule set (castling, en-passant, promotion, check, checkmate), a Raft-vs-Paxos explanation, zero-knowledge proofs, a geometry proof, a race-condition refactor, a blameless post-mortem, long-context recall. Core 27B handles far more of this than its size suggests. It isn't flawless, and the very hardest, most novel problems still favor the cloud — but the day-to-day ceiling is high.

The honest takeaway

If you came from Claude for everyday coding, writing, explaining, and translating, you would not feel a downgrade on Core 27B for most of what you do. The real differences live elsewhere: how fast it feels, where your code goes, and what it costs you to keep using it. That's the trade you're actually making — local privacy and one-time pricing against cloud speed and a higher ceiling on the hardest problems.

Frequently asked questions