Hardcore Reviews
Hardcore Reviews

Five ways to put AI on a phone, three licence traps

This review compares one front only: what forms an AI agent takes on a phone, not which vendor's model scores highest. Framing is fixed at the top: everything is a representative comparison assembled from official pages, repository READMEs and public reporting rather than independent benchmarking in a common environment, and exact numbers follow each vendor's live interfaces. It also states its division of labor against three existing articles on the site, the desktop computer-use agent comparison, the agent harness comparison covering the scaffolding that wraps a model as a different layer, and the phone-harness open-source piece, and limits itself to the phone side. The five contenders are Alibaba's Qwen Intelligence, a B2B stack for OEMs; Apple Intelligence with Siri AI, a system-level self-contained loop; OmniBot, an Android-native on-device open-source agent; OpenGUI, an Android phone GUI agent framework; and MobileGym, a browser-hosted phone simulation and evaluation platform. MobileGym is called out as a different layer, an evaluation and simulation base rather than an assistant, and is not compared as a peer assistant. Three licence red lines get their own section: OmniBot's segmented dual licence of AGPL v3 plus commercial that bars use by any organization; OpenGUI's Business Source License 1.1, source-visible but not OSI open source, with an Additional Use Grant of None limiting it to non-production use and commercial licensing needed for production, converting to Apache 2.0 on the 2030-04-29 change date; and MobileGym's split licence with Apache-2.0 code but CC BY-NC 4.0 non-commercial data. All three report NOASSERTION or inaccurate licence fields through the GitHub API, so the conclusions come from the LICENSE files. Cost is explicitly declared not directly comparable: Qwen at about 2.41 dollars per thousand tasks by vendor statement, Apple not billed separately but limited to expensive models with an unpriced paid tier, and the open-source options free as software but requiring your own models and devices with possible paid commercial licensing, never converted into a single ranked figure. Apple's framing is layered by source: public launch information spanning the September 14, 2026 release with iOS 27, English-only beta, five more languages in October, 16 languages including Simplified and Traditional Chinese, mainland China outside the launch, initial unavailability in the EU on iOS, iPadOS and watchOS but availability on macOS and visionOS, and daily usage limits on server-backed features with neither quotas nor paid-tier pricing published; the vendor table's description of Apple marked as source-side framing; and the report on the China-market Qwen backbone marked as media framing and distinguished from Qwen Intelligence.

Published September 23, 202610 min read
<!-- mobile-gui-agent-comparison-review | review | Five ways to put AI on a phone, three licence traps -->

Putting AI on a phone is at least five non-overlapping roads: a platform sold to handset makers, a first-party system assistant, an open-source app running locally, a cross-app automation framework, and a simulated phone in a browser. This piece does not rank model scores. It asks what forms a phone-side AI agent takes, and where the boundaries, openness, and licence red lines are.

This is a representative comparison: details come from official pages, repository READMEs, LICENSE texts, and public reporting, collected 2026-09-23, not an independent stress test under one identical setup. Official interfaces and docs are authoritative.

Three site articles are easy to confuse with this one. The Computer Use Agent comparison covers desktop agents (Claude, OpenAI, Mariner, cua, Agent-S, fara); this piece is phone-side only. The Agent Harness comparison covers model scaffolding (DSH, Claude Code, OpenCode, Codex), another layer. The phone-harness article covers an iPhone driven through a Mac harness, not re-dissected here.

Five offerings, five layers

#OfferingVendorFormWhere it runsLayer
1Qwen IntelligenceAlibaba Tongyi QianwenFull-stack AI solution for handset makers (model + platform + scenarios)Phone + backend / cloudB2B foundation
2Apple Intelligence / Siri AIAppleFirst-party system assistant (on-device model + cloud-side AFM)Device + cloudFirst-party closed loop
3OmniBotomnimind-ai (open source)On-device AI agent running directly on AndroidOn the Android deviceOpen-source on-device app
4OpenGUICore-MateAndroid phone GUI agent frameworkPhone + local backendSource-available framework
5MobileGymPurewhiter (academic)Browser-hosted phone simulation and evaluation platformBrowser simulationEvaluation / simulation substrate

The fifth row is the one most often misread: the first four are assistants or near-assistants, the fifth is not. MobileGym sits on a different layer: it does not play assistant, it measures the others. Pitting it against the first four is the common category error here.

The five, one by one

Alibaba Qwen Intelligence: a foundation sold to handset makers, not an app you download. A B2B full-stack offering that builds no hardware, selling handset makers "model + platform + scenario solutions"; it has landed with Honor. Three agents split the work: Mobile Planner for planning, 100-plus-step tasks, and dual-layer memory; Mobile-Use for cross-app operation on an "API first, GUI as fallback" path, with user confirmation for key decisions; Mobile Creative for images from colloquial requests.

Vendor-stated figures (vendor claims, not our re-test): 91.8% overall task accuracy; about USD 2.41 per thousand Planner calls; Mobile-Use 90% end-to-end success, 28.8% less time, 20% lower cost, 59.5 seconds per task; Creative's first image in 3 seconds; a security score 23 points above leading options. It opens four evaluation suites, but consumers only get Honor's shipped devices and a Qwen platform trial zone. Licence and pricing are undisclosed. Background: the Qwen Intelligence launch hotspot.

Apple Intelligence / Siri AI: a first-party closed loop with three hard constraints. Claims must be labelled by tier. Apple's public release information: launched 2026-09-14 with iOS 27 and its siblings, English only in Beta, with French, Japanese, Korean, Portuguese and Spanish in October; 16 languages including Simplified and Traditional Chinese. The regional red line matters most: mainland China is not in the first wave, pending Apple's completion of regulatory requirements, and the EU cannot use the features on iOS, iPadOS or watchOS initially, though macOS and visionOS can. Server-side features carry a daily usage limit with an undisclosed quota; a paid tier is coming with no announced price or date.

Two further framings must stay separate. Counterparty framing: the Qwen Intelligence table calls Apple "on-device in-house model, driven by Google Gemini". Press framing: a ZOL report on 2026-09-10 said Apple's initial release completed its in-country large-model filing in July 2026, supported by Alibaba's Qwen model. That is not Qwen Intelligence, the B2B foundation sold to handset makers.

OmniBot: an open-source on-device agent that truly runs on the Android phone. GitHub API snapshot, 2026-09-23: 2,007 stars, 144 forks, Dart, created 2026-03-18, last push the same day. It runs directly on the Android device (Kotlin plus Flutter, with a React + TypeScript WebUI). The surface is wide: Skills, an Alpine Linux environment, a browser, MCP, Android system tools, alarms, calendars, audio, memory, files, and a terminal. A Remote Codex bridge lets the phone use a signed-in computer's Codex via a QR code. The local service defaults to port 8899.

Unlike the site's phone-harness piece (iPhone via Mac harness), this is Android-native. Repo overview: OmniBot, the open-source on-device agent.

OpenGUI: an Android real-device GUI agent framework, source-available but not open source. GitHub API snapshot, 2026-09-23: 1,804 stars, 114 forks, TypeScript, created 2026-04-18, last push 2026-09-17. It lets an agent see, understand, and operate Android app interfaces on a real device. The shortest path is a DeepSeek Harness (DSH) plugin installed with one prompt to Codex; full-stack self-hosting is supported. It supports DSH 0.1.0-rc.7 through 0.1.1-rc.2 (default 0.1.1-rc.2), not 0.1.2-alpha.4; the macOS installer needs Node.js 22.19+.

The design is "plan before acting, summarise after execution": a supervisor keeps task state, the executor runs a screenshot-vision-action-callback loop, and a summariser closes with a structured result; planning and VLM execution can use different providers; Feishu, Telegram, Discord and REST dispatch are supported. Per the README: Android 11 (API 30) or newer, USB debugging and AccessibilityService permissions required; quality depends on model, app UI, network and task length; it is not an always-on assistant yet.

MobileGym: another layer, an evaluation and simulation substrate, not an assistant. Up front: it is not an AI assistant on a phone; it is the substrate used to quantify the others, not a same-kind comparison. Paper figures (arXiv 2605.26114, EMNLP 2026 main conference): 28 simulated apps, 416 parameterised task templates, deterministic sub-millisecond judges; 256 parallel instances on one server at roughly 400 MB of RAM each and about 3 seconds of cold start. Sim-to-Real: a GRPO run on Qwen3-VL-4B gains 42.8 points in simulation and keeps 95.1% of that on a real device, or 40.7 points on hardware. Paper and README figures, not our re-test. To set up an on-device agent yourself, see the phone-side agent setup SOP.

DimensionQwen IntelligenceApple IntelligenceOmniBotOpenGUIMobileGym
Cross-app pathAPI first + GUI fallbackDeep system integrationSystem tools + Skills + MCPAccessibilityService, GUI visionN/A (evaluation substrate)
Authorisation and safetyThree-layer controls; user confirmation for key decisionsOn-device processing + daily server-side limitUse constrained by licenceUSB debugging and accessibility permissionsSandbox, rollback, deterministic judging
OpennessCustomisable for vendors; four evaluation suitesClosed ecosystem, developer APIs onlyOpen source (AGPL v3, scoped)Source-available (BSL 1.1, not OSI)Code Apache-2.0, data CC BY-NC 4.0
Licence red lineUndisclosed; official statements governNot applicableCommercial use needs a licenceProduction/commercial use needs a licenceData non-commercial, academic only
Cost basisAbout USD 2.41 per thousand tasks (vendor)Not billed separately; high-end devices onlySoftware free; BYO model and deviceSoftware free; BYO model and deviceOpen source; BYO servers and compute

Three licence red lines

All three report NOASSERTION or an inaccurate value in the GitHub API licence field; conclusions below come from the LICENSE text itself.

Red line one: OmniBot uses a Segmented Dual Licence. Its LICENSE says the open-source side is GNU AGPL v3, free only for non-commercial use or personal, educational, or research purposes, and it explicitly "prohibits use by any organisation"; AGPL source obligations still apply. Commercial use, or avoiding AGPL's source-release obligation, needs a commercial licence signed with the maintainer. Even research by an organisation falls outside the free scope; calling it "AGPL open source" or "free for commercial use" is wrong.

Red line two: OpenGUI uses Business Source License 1.1 (BSL 1.1), source-visible but not OSI open source. You may use the source for non-production purposes, but production or commercial use, hosted services, and commercial integration need a licence from Core-Mate. For this version, Additional Use Grant: None (non-production only), and the Change Date is 2030-04-29, when it converts to Apache 2.0. In 2026 it is public source, not open source. The GitHub API reports NOASSERTION here.

Red line three: MobileGym splits code and data licences. Code (core source and mobilegym-rl/) is Apache License 2.0; data and content (mobilegym-data/, per-app data and assets, synthetic content, simulated UGC, icons) is CC BY-NC 4.0, non-commercial academic use only. The split is deliberate: code stays permissively reusable, content (including derived representations of third-party brands) stays research-scoped. Academic evaluation is fine; a commercial product crosses the line on the data side.

Cost: three billing logics, not directly comparable

Bases only; never one number.

  • Qwen Intelligence: about USD 2.41 per thousand Planner calls (vendor figure), a fraction of leading models; per-call pricing, rules unpublished.
  • Apple Intelligence: not billed separately but limited to high-end new devices; server-side features carry a daily usage limit and the paid tier has no announced price or date. Hardware-tied.
  • The open-source side (OmniBot, OpenGUI): software free, but you bring your own model API or compute, devices and backend; commercial use may add a licence fee. Self-funded infrastructure.
  • MobileGym: open source, but you bring your own servers and compute.

Four different bases; converting them into one number would be wrong.

Selection by scenario, and the verdict

Your needPickWhy
A vendor-customised foundationQwen IntelligenceA one-stop AI foundation for handset makers, customisable
A first-party system assistantApple IntelligenceDeep system integration, but bound by region and device
A controllable on-device open optionOmniBotRuns directly on Android; commercial use must clear the licence first
A cross-app automation frameworkOpenGUIA real-device GUI framework, usable as a DSH plugin; production needs a commercial licence
Large-scale evaluationMobileGymAnother layer: a simulation substrate that buys reproducible, quantified results

Three conclusions. First, these five are not pick-one-of-five; they are three non-overlapping tracks plus one evaluation layer: a foundation platform, device-side apps and frameworks, and the simulation substrate that quantifies them. Second, openness is tiered, not binary: Apple is closed, Qwen is vendor-customisable, OpenGUI is source-available but not open source, and OmniBot and MobileGym are open source with red lines. Treating "source-available" as "open source", or "open source" as "free for commercial use", are the two most common traps. Third, licence and cost must be read together: whether "free" includes commercial rights decides if it can enter production.

FAQ

Q1: Of these five, which is most worth using today? A1: It depends on region and device. Apple's option is limited because mainland China is not in the first wave; Qwen Intelligence only reaches consumers through Honor devices and a Qwen platform trial zone. What you can try today is OmniBot and OpenGUI, at the cost of your own hardware and licence constraints.

Q2: Why is MobileGym not counted as a phone assistant? A2: Because it is a browser-hosted simulation and evaluation platform, not an assistant that runs on a real device to do work for the user. It provides a programmable, rollback-capable, deterministically judged environment. It is the layer that measures assistants, not one of them.

Q3: OpenGUI calls itself "open source", so why can it not be used commercially? A3: Because it uses BSL 1.1, source-visible but not an OSI-approved open-source licence. Additional Use Grant: None grants non-production use only; production and commercial use need a separate commercial licence until the Change Date of 2030-04-29, when it becomes Apache 2.0. The LICENSE text governs, not the GitHub API's NOASSERTION field.

Q4: Is a self-hosted open-source option always cheaper? A4: No. OmniBot and OpenGUI are free as software, but you supply model APIs or compute, devices and backend, and commercial use may add a licence fee. Qwen is priced per call (vendor figure, about USD 2.41 per thousand tasks). Different bases should not be reduced to one number.

Q5: Are Apple Intelligence and the Qwen foundation a China-market iPhone may use the same thing? A5: No. The ZOL report of 2026-09-10 (press framing) said Apple's initial release completed its in-country large-model filing, supported by Alibaba's Qwen model: the foundation of an in-country system. Qwen Intelligence is the B2B full-stack offering Alibaba sells to handset makers. Both carry the name "Qwen", but they are not the same thing.


Sources

  • Official pages and READMEs: Qwen Intelligence official comparison table (counterparty framing), Apple Intelligence / Siri AI release information, omnimind-ai/OmniBot, Core-Mate/OpenGUI, Purewhiter/mobilegym
  • LICENSE texts: OmniBot's LICENSE (AGPL v3 plus commercial dual licence), OpenGUI's LICENSE (BSL 1.1, Change Date 2030-04-29), MobileGym's LICENSE and LICENSE-DATA (Apache-2.0 / CC BY-NC 4.0)
  • GitHub API snapshots (2026-09-23): OmniBot 2,007 stars; OpenGUI 1,804; MobileGym 798. All three licence fields read NOASSERTION or are inaccurate; the LICENSE text governs
  • Paper: MobileGym, arXiv 2605.26114, EMNLP 2026 main conference
  • Public reporting: ZOL, 2026-09-10 (press framing); Apple launch information relayed by Mac Observer, AI Insiders, ET Now and HandWiki

Collected 2026-09-23; star counts are same-day snapshots; features, licences, and prices follow the official sources. Related reading: Qwen Intelligence launch hotspot | OmniBot, the open-source on-device agent | Phone-side agent setup SOP | Agent Harness comparison | Computer Use Agent comparison

This article is AI-assisted and human-edited. Last updated: 2026-09-23

FAQ

Of these five, which is most worth using today?
It depends on region and device. Apple's option is limited because mainland China is not in the first wave; Qwen Intelligence only reaches consumers through Honor devices and a Qwen platform trial zone. What you can try today is OmniBot and OpenGUI, at the cost of your own hardware and licence constraints.
Why is MobileGym not counted as a phone assistant?
Because it is a browser-hosted simulation and evaluation platform, not an assistant that runs on a real device to do work for the user. It provides a programmable, rollback-capable, deterministically judged environment. It is the layer that measures assistants, not one of them.
OpenGUI calls itself "open source", so why can it not be used commercially?
Because it uses BSL 1.1, source-visible but not an OSI-approved open-source licence. `Additional Use Grant: None` grants non-production use only; production and commercial use need a separate commercial licence until the Change Date of 2030-04-29, when it becomes Apache 2.0. The LICENSE text governs, not the GitHub API's NOASSERTION field.
Is a self-hosted open-source option always cheaper?
No. OmniBot and OpenGUI are free as software, but you supply model APIs or compute, devices and backend, and commercial use may add a licence fee. Qwen is priced per call (vendor figure, about USD 2.41 per thousand tasks). Different bases should not be reduced to one number.
Are Apple Intelligence and the Qwen foundation a China-market iPhone may use the same thing?
No. The ZOL report of 2026-09-10 (press framing) said Apple's initial release completed its in-country large-model filing, supported by Alibaba's Qwen model: the foundation of an in-country system. Qwen Intelligence is the B2B full-stack offering Alibaba sells to handset makers. Both carry the name "Qwen", but they are not the same thing.

Related

Hardcore Reviews

TrnRT, Kindi, Baseten: One License Forbids Managed Service

Three inference engine repositories, TrnRT, Kindi and Baseten, on one comparison table, with license terms and hardware bets as the axis rather than a buying recommendation or benchmark ranking. Apache-2.0, Elastic License 2.0 and AGPL-3.0 coexist; one LICENSE says verbatim that the software may not be provided to third parties as a hosted or managed service, and GitHub reports NOASSERTION. Hardware claims from NKI megakernels on Trainium2 to GB300, B300, Quadro K420 and RTX 3080 are labelled as README self-reports. Stars are cumulative totals, with no growth score or single-day delta.

Nov 16, 20269 min read
Hardcore Reviews

qwen-code Preview, OpenBot, free-claude-code: Three Compared

Sets the qwen-code v1.0.0-preview beside OpenBot v1.0.2 and the rolling v6.x of free-claude-code in one table, comparing version states and access boundaries rather than picking a winner. Five rows cover models and providers, where AG-UI and MCP attach, permissions and approvals, stability during Preview, and license points. The free-claude-code column states AGPL-3.0-only, the NOASSERTION value from GitHub, the README disclaimer that it is not affiliated with Anthropic, and that provider terms govern integrations. Star counts are cumulative totals.

Nov 15, 20269 min read