Outlier vs Qwen Chat: the same model family, two delivery models

Corrected 2026-09-29: an earlier version listed Outlier's Vision tier as Qwen3.6-35B-A3B (19.0 GB), the model it replaced; Vision 3.8 is built on Qwen3.8-27B, a 15.5 GB download.

I'm Matt Kerr, and I built Outlier, so treat this as a comparison by an interested party. Every claim about Alibaba's product is tied to a source you can check. This one's odd to write, because Outlier runs Qwen weights: five of its six tiers are Qwen models (Quick is Gemma 4). One family, two very different deliveries.

What Qwen Chat actually is

Qwen Chat is Alibaba's consumer AI assistant, at chat.qwen.ai. The web app's page title now reads just "Qwen," one iOS listing is still titled "Qwen Studio," and qwen.ai still serves the old /qwenchat route. It's operated by Alibaba Cloud (Singapore) Private Limited.

The app's public model endpoint lists exactly three models: Qwen3.8-Max, Qwen3.7-Plus and Qwen3.8-Omni-Flash. All three are flagged for vision, documents, video, audio and thinking, and each offers search; Qwen3.7-Plus and Qwen3.8-Max also carry the image-generation, deep research and slides modes.

There's no consumer subscription. The config reports subscriptions and payment/quota disabled and anonymous use enabled, so you can open the web app and start typing without an account; sign-in is Google, GitHub or Apple OAuth. Native clients cover macOS, Windows, iOS and Android.

The part that makes this comparison strange: same family, different weights

Outlier's tiers are built on Qwen3.5-4B, Qwen3.5-9B, Qwen3.6-27B, Qwen3.8-27B and Qwen3.5-397B-A17B, all Apache-2.0. (Quick is a Gemma 4 model, also Apache 2.0; see the model card.) Those are Alibaba's published open models. You download them once and run them on your machine: Nano, Lite, Quick and Core from Outlier's own Hugging Face repos, Plus and Vision 3.8 from public community conversions.

Here's the honest wrinkle: you can't run what Qwen Chat is serving on a Mac. Qwen3.8-Max is the hosted version of Qwen3.8-2.4T-A95B, which Alibaba published as open weights in August 2026 under its own license rather than Apache-2.0; its model card calls it the first Max-class model Qwen has released openly. I found no published weights for Qwen3.7-Plus. Qwen3.6-35B-A3B landed 15 April 2026 and Qwen3.6-27B on 21 April 2026: both Apache-2.0. Qwen3.6-27B is the base of Outlier's Core tier; the 35B-A3B was the base of Vision 3.6, which Outlier has retired (Vision 3.8 is built on Qwen3.8-27B). Alibaba still ships genuinely open models, which I'm grateful for, but the chat product runs a tier above anything I can put on a laptop.

The flip side is the escape hatch: if you later decide you don't want your work on someone else's servers, you can move to Qwen weights you control without changing families. That's rare among cloud assistants, and it's most of why Outlier exists.

Where Qwen Chat genuinely wins

These aren't close calls. If any of them describes your work, use Qwen Chat.

• Free, no account, nothing on disk. Outlier's free tiers cost nothing either, but you download a 2.37 GB model first.
• It runs models no Mac can run. The Max tier is datacenter-scale: Qwen3.8-Max is built on a 2.4-trillion-parameter model. Outlier's largest tier, Plus 397B, is a 209 GB download needing a 64 GB Mac, at 1.59 tokens/second on an M1 Ultra.
• A far stronger research stack. Search is a native capability on all three models Qwen Chat serves, and deep research is a first-party mode alongside it. Outlier does have opt-in web search with citations (DuckDuckGo, Mojeek, Wikipedia) and its own Deep Research mode, but both are Pro-only, both borrow third-party engines, and a local model reads the pages, so a wide sweep is slower and shallower than the same sweep run on a datacenter model. On the free tiers there's no web access at all.
• Image generation, image editing and text-to-video. Outlier reads images (the Vision tier) and transcribes your speech on Pro, and any reply can be read back in the system voice, but it generates no images and no video.
• Multimodal input local hardware can't touch. Video up to 2 GB and one hour, audio up to three hours, 20 MB documents. No Mac reasons over an hour of video on battery.
• It's on your phone. iOS and Android clients plus Windows and web, and it's fine on an 8 GB Mac. Outlier is Apple Silicon Mac only: no iPhone, no Windows.
• A real productivity surface. MCP tools, artifacts, web-dev mode, slides, projects (10 × 5 files), memory, 17 languages across 19 locales, and Qwen's strength in Chinese-language work.
• Good output terms. You keep rights in your prompts, and Qwen assigns you its interest in the outputs.

Cloud Qwen vs local Qwen, line by line

*On 2026-10-19 an App Store search returned no Qwen app for the US or German storefronts, while Singapore, UK and Japan returned the app. US iPhone users are on web or desktop.

Free to use, paid to build on

For anyone using the chat app: $0. There's no consumer tier to buy. That's hard to compete with, and I won't pretend otherwise.

The money is on the developer side. On Alibaba Cloud Model Studio, international list price for qwen3.8-max is $2 per million input tokens and $6 per million output, qwen3.7-max is $2.50 in / $7.50 out, and qwen3.7-plus is $0.40 in / $1.60 out at the 0–256K tier. Free API access is a one-million-token trial per model, valid 90 days, and only in the Singapore region. Token Plan (Team Edition) seats list at $30 to $200 a seat per month, $360 to $2,400 a seat per year, with limited-time discounts on the two lower tiers.

Outlier Pro covers all six tiers; the Outlier Lifetime Licence is $49. Nano and Lite don't expire and don't ask for an email.

The two clauses that decide this for NDA work

Everything runs cloud-side, so everything is uploaded. The Privacy Policy says Qwen collects "your prompts and other content you upload, such as text, files, images, audio and videos," plus device ID, IP address and interaction data. For voice and video it notes this "involves the processing of data that may be considered biometric identifiers."

Two clauses deserve a slow read. First, the Terms say you authorise use and storage of non-personal user content to develop and improve Alibaba's machine-learning models. Training on your content is the default; the training-data summary says users can opt out "by submitting their requests to us": a request to a company, not a switch in the app. Second, the Terms require a perpetual, irrevocable, sub-licensable, worldwide licence over user content and deem all of it "non-confidential and non-proprietary." That rules the product out for NDA-bound or client-confidential material.

Jurisdiction matters too. The Privacy Policy says data may be transferred out of and processed outside the jurisdiction you're in, and that certain group entities in Singapore and Mainland China get limited remote access to it. It describes encryption in transit and at rest and a security programme following "a standard industry framework," but names no certification. Terms are 18+.

Outlier's version of this section is short because there isn't much to say. Inference happens on your Mac: no account, no conversation telemetry, no opt-out to find.

Which Qwen should you run?

Pick Qwen Chat if you want large-model answers for free, you want agentic deep research rather than a single search pass, you want to generate images or video, you're on a phone or a Windows machine, or what you're pasting in isn't sensitive. For general knowledge work and drafting, free plus multimodal plus search is a strong package.

Pick Outlier if the content is the problem: client files, source code, medical notes, legal drafts, anything under NDA. The "non-confidential and non-proprietary" clause and the default training use make Qwen Chat a poor fit there, and model quality doesn't fix it. Same if you need it to work on a plane, or if you want to still be running this exact model in five years.

Use both: that's what I do. Qwen Chat for wide public research, generated images and anything involving a video file; Outlier for work I'd be uncomfortable pasting anywhere. On a blind slice of SWE-bench Verified Core 27B measured 46%, and I published a 54-prompt side-by-side with Claude Opus so you can compare the answers yourself. Enough for most everyday work, and still not a frontier cloud model.