Outlier vs Zed AI: on-device Mac models vs a cloud-first code editor

Disclosure up front: I'm Matt Kerr, I built Outlier, and I sell it. Read this like any comparison written by an interested party, and check the sources at the bottom. I've worked to get Zed right, partly because it is better than Outlier on several axes I care about.

What Zed's AI actually is

Zed is a desktop editor written in Rust by Zed Industries, the team behind Atom and Tree-sitter. It's open source under GPL-3.0-or-later (Apache-2.0 where marked), runs natively on macOS, Linux and Windows, and ships a CLI.

The AI arrives in three shapes: an agent panel that reads and edits across your project, inline edit prediction as you type, and external agents over the ACP protocol — Claude Agent, Codex, OpenCode — plus MCP servers. As of mid-2026 Zed's docs describe a desktop application; I found no official mobile app or web version.

Where each one lives

Zed's AI lives inside the thing you write code in: it sees your cursor, buffer, diagnostics and repo, and edits stream across files as the agent works.

Outlier is a standalone Mac app — Apple Silicon only, M1 or later, macOS 12+. No editor extension, no inline completion. Its Code Agent reads a project folder and edits files behind approval gates, but it does that in its own window rather than in your buffer, which is more app-switching in a tight edit loop. What you get instead runs with Wi-Fi off and never asks for a sign-in.

Which model is doing the work

In Zed the heavy reasoning comes from closed-weight Claude, GPT and Gemini models, with the most capable tiers gated to Pro and Business. The one open-weight piece is Zeta 2, the edit-prediction model — fine-tuned from ByteDance's Seed Coder 8B, Apache-2.0, and open-weight on Hugging Face. It replaced the original Qwen2.5-Coder-7B-based Zeta as Zed's default in March 2026. There are five access paths: Zed-hosted billing, your own keys, existing ChatGPT/Claude/Copilot subscriptions, gateways like OpenRouter and Bedrock, and local servers running Ollama, LM Studio or llama.cpp.

Outlier ships seven tiers, all on your Mac: Nano 4B (2.37 GB, 6 GB RAM) through Code 27B (15.13 GB, 24 GB RAM) up to Plus 397B-a17b, which streams experts from SSD and wants 64 GB. Six are Apache 2.0; Quick is under the Gemma Terms of Use.

Does your code leave the machine?

Zed is hybrid but cloud-first by default: using AI features sends prompts and code context to the provider. Zed doesn't retain that data, and its hosted providers operate under zero-data-retention agreements — with one documented carve-out: Anthropic retains data for its designated Covered Models for at least 30 days for trust-and-safety review rather than training. Your input isn't used for training by default, and edit-prediction collection is opt-in behind three conditions, restricted to open-source-licensed projects, and admin-controlled on Business. A conservative posture, honestly documented.

Even so, default edit prediction needs a sign-in and a cloud round-trip at keystroke granularity. The local path is documented, but it's a configuration you opt into rather than the architecture — local edit prediction drops External Agents and Terminal Threads, and Ollama's context defaults to 4096 tokens unless you raise max_tokens.

Outlier has no cloud path on by default: no account, no upload, works with the network off. Web research and external API keys exist as opt-in paths on Pro, and turning one on sends that specific request off the device by design — but nothing routes outward until you ask it to. The cost is real: with those switched off there's no live retrieval and no cloud sync, so the model works from what's on your disk.

What it costs

Zed's Personal plan is $0 — "$0 forever" on the pricing page — with 2,000 accepted edit predictions a month and unlimited AI use on your own keys. Pro is $10/month including $5 of tokens; past that it's API list price plus 10%, billed at month end or per additional $10 incurred, whichever comes first. Business is $30/seat/month with no seat minimum, no free trial and no bundled credits. The two-week Pro trial carries $9 of credits and excludes Anthropic's Opus models. No annual discount is published.

Credit where it's due: in September 2025 Zed moved from 500 prompts a month to token billing and cut Pro from $9 to $10. Worth checking — after the March 2, 2026 terms overhaul, the AI service requires users to be 18 or older, with binding arbitration and a class action waiver (30-day opt-out) under Delaware law.

Outlier's free tier is Nano and Lite, no account. Pro covers all seven tiers, Founders Lifetime at $249. Nothing is metered, because there's no server to meter.

Agentic ability and repo-scale context

Here are my coding numbers, and they don't all flatter me. On a blind slice of SWE-bench Verified the local 27B measured ~45% — 18 of 40. Code 27B scores 0.866 on HumanEval. Those are my own measurements, not leaderboard submissions, and they sit below the strongest cloud coding agents. Core 27B did match Claude Opus on 98.9% of rubric checks across a 54-prompt head-to-head — but that's answer quality, not multi-step agentic repair.

On repo-scale context Zed's advantage is straightforward: the agent holds your project plus a frontier context window, with ACP agents and MCP servers pulling in more. Outlier's context is bounded by the RAM in your Mac, and Plus 397B runs at 1.59 tok/s on an M1 Ultra — thorough, not quick. For "read forty files and refactor the call graph," cloud wins.

Zed default, Zed local, and Outlier on six axes

Where Zed's AI genuinely wins

• Frontier reasoning on demand. ~45% on a blind SWE-bench Verified slice is below what the best cloud coding agents deliver. On hard multi-step engineering, Zed driving Claude or GPT is the better tool.
• It's in your editor. Cursor context, live diagnostics, streaming multi-file edits. Outlier can't do that.
• Actually open source. GPL-3.0-or-later for the whole editor, so you can read exactly what the AI integration sends and when. Few tools allow that.
• The free tier is usable, not a demo. Bring your own Anthropic or OpenAI key and you never pay Zed anything.
• Cross-platform. Linux and Windows are first-class. Outlier is Apple Silicon Mac only — no Windows, no Linux, no mobile.
• Legible billing. List price plus 10% is forecastable from published provider rates.
• Vendor-agnostic. Five access paths plus ACP and MCP, so you're not welded to one provider.
• Honest docs. Zed's May 2026 post on local AI says the hardware for frontier models at good speed is "out of reach for consumers," and flags smaller context windows and lower throughput. I run a local-AI company and think that's fair.

Editor or standalone: which fits your day

These two aren't really competing for the same hour, so the useful question isn't which is better. It's what your code is allowed to touch.

If the answer is "anything," Zed is the better default: an agent working across a large repository, the strongest available reasoning on hard bugs, a client whose source you can audit, and the only one of the two that runs on Linux or Windows. It costs nothing if you already pay for Claude or OpenAI keys.

If the answer is "not a third party" — client repos under NDA, regulated systems, unreleased product, security work — Outlier is the one that guarantees it by architecture instead of by contract. The same holds if you work offline a lot, dislike metered billing, or don't want another account.

Most developers land on both, and that's what I'd actually suggest: Zed as the editor with your own keys or a local model, Outlier for the slice of work that shouldn't go over the wire.