Outlier vs Windsurf (Codeium): local coding AI vs a cloud agent IDE

I'm Matt Kerr and I build Outlier, so read this as a comparison written by an interested party. Every claim about the other product is sourced to its own docs, and my numbers are here even where they're unflattering.

Read this first: "Windsurf" is now Devin Desktop

Windsurf is a retired name. Codeium became Windsurf, Cognition acquired it, and on June 2, 2026 it became Devin Desktop; windsurf.com now 308-redirects to devin.ai/desktop.

Plans, pricing and extensions carried over. The agent is mid-transition: Cascade, the agent that defined Windsurf, is still the shipping default, while Cognition's next-generation Devin Local agent — which its docs say takes up to 30% fewer tokens than Cascade for the same task — is in preview and not yet at parity: it doesn't persist memories between sessions, doesn't support workflows, and isn't available in the JetBrains plugin. Cognition says it's meant to eventually replace Cascade. Three product identities in two years is a real risk if you're standardizing a team.

Where each one lives

Outlier is a native macOS app for Apple Silicon, M1 or later, macOS 12+. That's the whole surface area: no Windows or Linux build, no mobile app, no web version, no extension.

Devin Desktop is the opposite shape — a full IDE with the agent inside it, built for macOS, Windows and Linux, plus a command-line agent and a JetBrains plugin. If you want something that edits your working tree and iterates until the tests pass, that's what it was made for. Outlier isn't an editor.

Which model is doing the work

Outlier ships seven tiers that download once and run on your GPU, from Nano 4B to Plus 397B-a17b, which pages MoE experts off the SSD. Six are Apache 2.0, the weights are on Hugging Face, and on an M1 Ultra Nano runs about 71.7 tok/s and Core 27B about 20.7.

Devin Desktop routes to a large hosted catalog — OpenAI, Anthropic, Google, DeepSeek, Moonshot, Zhipu, Minimax, xAI. Cognition doesn't publish a headline number, so I counted the model table at docs.devin.ai/desktop/models on 2026-07-23: about 170 entries on Pro and about 175 on Enterprise (169 and 174 distinct labels, since a couple appear twice).

On top of that catalog sit Cognition's own SWE models, and their claims are per-model rather than family-wide. SWE-1.5 was announced as a frontier-size model with hundreds of billions of parameters, served on Cerebras at up to 950 tok/s. SWE-1.7, the current one, is quoted at 1000 TPS via Cerebras and ships with no published parameter count; the models doc now calls SWE-1.5 "our previous frontier agentic coding model."

All of it is vendor-hosted. "Devin Local" is a local agent harness with access to your files and tools, not local weights. The ACP docs list five example pluggable agents — Codex CLI, Claude Agent, OpenCode, Junie and Gemini CLI — and say any agent that speaks ACP can run inside Devin Desktop. What that establishes is narrower than "no local option": no Devin-supplied model runs on your machine, and Cognition documents no local runtime. Since OpenCode is open source, you could in principle point an ACP agent at local weights yourself — and the same doc tells you the cost of doing it, that Devin Desktop's privacy policy and legal terms no longer apply and billing goes directly to the third party. Out of the box, though: no network, no model.

Does your code leave the machine?

Outlier: no. No account, no round trip, no prompt telemetry. Turn Wi-Fi off and it keeps working, and there's no training opt-out to hunt for because there's nothing to opt out of.

Devin Desktop: yes for inference, and it depends for indexing. Inference is unconditional — every prompt and every file the agent reads goes to a hosted model.

Indexing is the part usually described wrongly, including by me before I checked. On the individual plans this page compares — Free, Pro and Max — the index is built on your machine: Cognition's context docs say the entire local codebase is indexed on device, and list indexing of remote repositories as something available additionally to Teams and Enterprise users. The Remote Indexing doc is blunter still: the feature "is only available in the Devin Desktop Plugins for Enterprise plans," and describes organizations on Teams and Enterprise plans cloning repositories into Cognition's indexing service to build embeddings on a single-tenant instance. On that remote path only, code and snippets are deleted after embedding, assuming "Store Snippets" is unchecked.

So if you're an individual on Pro, your repository is not being cloned to Cognition's servers. The exposure is real but smaller than the usual telling: the snippets your local index retrieves are attached to every request that goes to the hosted model, so the code still travels — a retrieved slice at a time, rather than a whole-repo copy sitting in someone else's storage.

Then training. Cognition's Platform Terms of Service, effective June 30, 2026, say it may use Customer Data for model training. Paid subscribers can opt out on the Data Controls page, which also enables Zero Data Retention; the free tier has no documented opt-out. One caveat: Cognition's enterprise security page says it does not train on customer data by default, contradicting those consumer terms — that page appears to cover enterprise deployments only.

What you pay

Devin Desktop: Free at $0 with a light agent quota and unlimited Tab completions; Pro at $9/mo, adding frontier models from OpenAI, Anthropic and Google plus cloud agents; Max at $249/mo for higher quotas; Teams at $80/mo base plus $40 per seat. Yearly terms aren't published, so I can't give an honest annual figure.

You also can't compute what $9 buys in advance: quota numbers aren't published, only that your daily quota exceeds a seventh of your weekly one, and overflow is billed at API list prices. Credits became quotas in March 2026, with existing subscribers grandfathered indefinitely.

Outlier: the free tier is Nano and Lite, no account, no cap. Pro covers all seven tiers; Founders Lifetime is $249. No quotas — the compute is yours.

Outlier vs Devin Desktop at a glance

Agentic ability and repo context: my honest receipts

This is where Outlier is behind, and I'd rather say so up front. On a blind slice of SWE-bench Verified the local 27B measured about 45% — 18 of 40: my own harness, not a leaderboard submission, published because it's real, not because it flatters me. Core 27B scores 0.866 on HumanEval. Against a hosted agent with hundreds of billions of parameters, that's a gap: on a long refactor with cross-module coupling, the cloud agent lands changes a 27B on your laptop won't.

Repo-scale context is the second gap, and it's a gap on both paths: Devin Desktop's local index on Free/Pro/Max is backed by a hosted model with a far larger context budget than a 27B holds in local RAM, and on Teams and Enterprise it can index whole remote repositories in the cloud on top of that. On a two-million-line monorepo neither difference is subtle. Throughput is the third — SWE-1.7 at 1000 TPS on Cerebras against Core 27B's 20.7 tok/s on an M1 Ultra.

Where local holds up: reading unfamiliar code, writing a function, explaining a stack trace, refactoring one file. Core 27B matched Claude Opus on 98.9% of rubric checks in a 54-prompt head-to-head — but that's general capability, not agentic coding.

Where Windsurf (Devin Desktop) genuinely wins

• Breadth, scale and speed no local app can approach. Roughly 170 models on Pro across eight vendors, switchable per request, so a bad answer costs you a dropdown rather than a re-download. A frontier-size model — SWE-1.5 was announced at hundreds of billions of parameters — can't sit resident on a 16 GB laptop at all, and SWE-1.7 at 1000 TPS on Cerebras is orders of magnitude past consumer Apple silicon.
• Cloud agents leave your machine alone. Long refactors run in remote VMs — no RAM, GPU or battery cost, and they survive a closed lid.
• Cross-platform. macOS, Windows, Linux, JetBrains, CLI. Outlier is Mac-only, which serves part of a mixed team at best.
• Deep agent tooling, with an escape hatch. MCP servers, skills, hooks, workflows, AGENTS.md rules, worktrees and subagents — plus ACP, which drives Claude Agent, Codex CLI or Gemini CLI from the same editor. Cognition built the door out of their own product.
• Compliance posture. SOC 2 Type II, a FedRAMP documentation track, SSO, SCIM, RBAC, customer-managed keys, dedicated deployment, and Zero Data Retention as one toggle on any paid plan. Outlier has none of it, having no server to certify.

Which one fits your repo

Pick Devin Desktop if you want an agent driving multi-file changes across a large repo; you need capability or throughput your hardware can't produce; your team is on Windows or Linux; procurement wants SSO and a SOC 2 report; or you want agents running while your laptop stays free.

Pick Outlier if your code sits under an NDA or a rule that makes sending source to a third-party model a non-starter — note that this is about inference, not indexing, since an individual Devin Desktop plan indexes locally but still ships retrieved snippets with every request; you want a structural guarantee rather than a toggle you pay to enable; you work offline; or you want a flat one-time cost instead of usage-based overflow billing.

Use both, what I'd tell most developers: keep the sensitive and offline work local, and send the gnarly refactor to a cloud agent with a far bigger model behind it.