Outlier vs the DeepSeek app: local Mac AI vs a free cloud chat service

Disclosure first: I build Outlier, so I have an obvious stake in one of these two products. Every DeepSeek claim below is tied to DeepSeek's own documentation, or marked when it comes from press reporting.

What DeepSeek actually is

DeepSeek is a cloud AI assistant from Hangzhou DeepSeek Artificial Intelligence Co., Ltd. Its site lists DeepSeek Web (chat.deepseek.com), the iOS and Android app, the DeepSeek Harness desktop app, and the API Platform.

The iOS listing, "DeepSeek - AI Assistant," is listed as Free, with no In-App Purchases block, needs iOS 15.0 or later, and supports iPhone, iPad and iPod touch only. DeepSeek's downloads page also offers DeepSeek Harness for desktop, for macOS (Apple silicon, macOS 13 or later) and Windows. Harness is DeepSeek's open-source agent app ("work in the background, edit local files"). I haven't tested it, and the pages I read don't say whether its model runs locally or in DeepSeek's cloud, so this page compares the chat app and the API.

Underneath is DeepSeek-V4, announced April 24, 2026 as live and open-sourced: V4-Pro at 1.6T total / 49B active parameters, V4-Flash at 284B / 13B. On September 10, 2026 it announced V4.1-Flash, a 552B-parameter MoE that replaced V4-Flash; the API now serves it as deepseek-flash. DeepSeek says 1M context is the default across all official services, with 384K max output. The weights are on Hugging Face, MIT-licensed.

Where DeepSeek genuinely wins

Several of these gaps are structural, and I can't close them.

• It's free. Not free-tier-with-a-paywall: the App Store listing is Free, with no In-App Purchases block on it. Outlier's paid tier competes against $0.
• It runs models no Mac can. 1.6T total parameters won't fit in 64 GB of unified memory at any quantization I'd ship, and a million tokens of context is the default on every official surface. No Mac holds that much KV cache.
• The weights are genuinely open. MIT-licensed and downloadable in FP8-Mixed and FP4+FP8-Mixed, with commercial use, modification and fine-tuning uncapped. Distrust the hosted service and you can self-host the same family. No closed vendor offers that.
• Independently measured, not self-claimed. NIST's Center for AI Standards and Innovation put V4-Pro at an IRT-estimated Elo of 800 against GPT-5.4 mini's 749 across nine benchmarks, and more cost efficient on 5 of 7.
• The API is cheap and drop-in. deepseek-flash costs $0.15 per million input tokens on a cache miss and $0.60 output off-peak; V4-Pro, $0.66 and $1.98. Peak rates are double. Both are OpenAI-compatible, with an Anthropic-compatible endpoint too.
• Generous downstream rights. Outputs are assigned to you, and the Open Platform Terms explicitly permit applying inputs and outputs to "derivative product development, training other models (such as model distillation)." Most vendors forbid precisely that.
• No hardware floor, and it's in your pocket. Nothing to download or manage, history syncs across devices, same behavior on an 8 GB laptop or a borrowed browser.

Outlier has no image generation, no mobile app, no hands-free voice mode, no Windows or Linux build, no cloud sync, and web search exists only on Pro, opt-in per conversation and narrower than DeepSeek's. If any of those matter to you, the comparison is over.

What Outlier is

Outlier is a native macOS app for Apple Silicon (M1 or later, macOS 15+). The model runs on your machine: no account, no usage cap, and it works with Wi-Fi off. Six tiers ship, from Nano 4B (2.37 GB, 6 GB RAM, 71.7 tok/s on an M1 Ultra) through Core 27B up to Plus 397B-a17b, which pages MoE experts off the SSD at roughly 14 GB peak RSS on a 64 GB Mac, at 1.59 tok/s. That last figure is the honest shape of the ceiling: the flagship tier needs a 64 GB machine and reads back at a walking pace. Weights are published.

On quality: on a blind slice of SWE-bench Verified the local 27B measured 46% (23 of 50, build 1.11.757, 2026-08-09), a real number and not a frontier one. We also published a 54-prompt side-by-side with Claude Opus so you can compare the answers yourself; cloud models are still faster and stronger on the hardest work.

Side-by-side

What each one costs over a year

The consumer chat is the easy column: as far as any primary source shows, $0. The App Store listing is Free, with no In-App Purchases block on it, and DeepSeek publishes no consumer subscription page. Third-party trackers describe web search, DeepThink mode, file upload and history sync as included free (reported, not documented).

The API is metered. A million input plus a million output tokens at cache-miss rates comes to about $0.75 on deepseek-flash and $2.64 on V4-Pro at off-peak prices, double that at peak, with cache hits dropping input to $0.003 and $0.022 per million off-peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Limits are per-account concurrency: 2500 connections for Flash, 500 for Pro. The model surface moves on a short clock: the pricing page now lists deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813), and the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but route to the Flash model.

Outlier: Nano and Lite are free forever, no account. Pro covers all six tiers; the Outlier Lifetime Licence is $49. Against a free product, that's a price you pay for locality, not capability per dollar, which DeepSeek wins outright.

What happens to your data

Everything DeepSeek does is remote. Its privacy policy, updated February 10, 2026, says the company "directly collect[s], process[es] and store[s] your Personal Data in People's Republic of China." What leaves your machine: text and voice input, prompts, uploaded files, photos, feedback and chat history, plus IP address, device identifiers, cookies, OS and approximate location.

Inputs are used for training by default. The stated purpose includes training and improving DeepSeek's machine learning models, and the policy discloses sharing with corporate-group entities for "foundation model training and optimization." The policy also grants "the right to opt-out of using your Personal Data for training our models or optimizing our technologies," without naming how you exercise it. In the product, the control is a toggle: Settings › Data controls › "Improve the model for everyone," on both web and mobile. Turning it off is forward-looking only. Data you shared before you flipped it stays in the training pool unless you email privacy@deepseek.com naming the specific chats you want purged. Worth knowing that the toggle appears in third-party guides rather than in DeepSeek's own privacy policy, so the documented commitment and the shipped control aren't the same artifact. The Open Platform Terms contain no clause excluding API inputs from training, and no retention policy at all.

Credit where it's due: the policy states in capitals that DeepSeek does not run targeted advertising, sell personal data, or profile. Consumer terms are governed by mainland PRC law, disputes heard in Hangzhou.

A few more items, from press and public-record reporting rather than DeepSeek's documents. Wiz Research found an exposed, unauthenticated database in January 2025 holding over a million entries including chat messages and API keys; South Korea's data-protection commission later found undisclosed transfers of over a million users' data. Government devices in Italy, Taiwan, Australia, Canada, South Korea, the Czech Republic, Texas and the US Navy have been restricted from DeepSeek since early 2025. On censorship: PromptFoo's January 2025 test of DeepSeek-R1, reported by TechCrunch, found canned pro-government responses to roughly 85% of 1,360 China-sensitive prompts, though the researchers called the restrictions crude and easily jailbroken. I have not seen an equivalent published measurement for V4, and I know of no published test of what a self-hosted copy of the weights does.

Outlier's half is short: the model runs locally, nothing is transmitted, there's no account to attach a history to, and no opt-out because there's no collection.

Who should pick which

Pick DeepSeek if you need a million tokens of context, want frontier-adjacent capability at zero cost, work across a phone and several machines, don't own an Apple Silicon Mac, or need live web results. For developers, the API is cheap and compatible enough that there's little reason not to try it.

Pick Outlier if the content of your prompts is the problem: client files, patient notes, unreleased source, anything under NDA or anything a regulator won't let cross a border. Also pick it if you're somewhere DeepSeek is restricted on work devices, or you need it to work on a plane.

Run both and you get the honest best of it: Outlier as the local default for sensitive, offline, high-volume work, DeepSeek for million-token questions that need the live web. Just remember that anything you send there is training data until you go turn that toggle off, and that turning it off doesn't reach backwards.