Locally AI + Qwen: Your Free Private AI, Running Locally on Your Phone Without Internet

2026-03-05 | Product Hunt | Official Site | App Store

30-Second Quick Judgment

What is this app?: A completely free iOS/iPad/Mac app that lets you run open-source models like Qwen, Llama, Gemma, and DeepSeek locally on your iPhone. No cloud, no login, no subscription fees—all data stays on your device.

Is it worth watching?: Yes. Although it only has 12 votes on PH, developer Adrien Grondin's Twitter demo of Qwen 3.5 running on an iPhone 17 Pro hit 6,906 likes and 3.43 million views. This suggests the product's appeal far exceeds what PH data reflects. Free + Total Privacy + Offline availability is a "no-brainer" combo for many.

Three Key Questions

Is it relevant to me?

Who is the target user?:

• Privacy-conscious individuals (who don't want to send chats to OpenAI/Google servers).
• People who don't want to pay $20/month for ChatGPT Plus.
• Users who need AI assistance in no-network zones (planes, subways, overseas).
• AI enthusiasts/developers wanting to test open-source models on mobile.
• InfoSec professionals needing AI analysis without data leaks.

Is that me?: If you've ever hesitated for a second about where your ChatGPT data goes, you are the target. If you've ever wanted AI to edit text on a plane without Wi-Fi, you are the target.

When would I use it?:

• Planes/Subways/Remote areas → Offline AI assistant for emails, copy editing, and translation.
• Handling sensitive files → Contract reviews, financial drafts; data never leaves the device.
• Casual chatting → A free AI companion.
• Siri + Shortcuts automation → Drive Apple ecosystem workflows with local models.
• When NOT to use it → If you only care about the strongest models (GPT-4o/Claude Opus) and aren't sensitive to privacy.

Is it useful to me?

ROI Judgment: Zero cost to try; it's a win just to have it installed. Treat it as a "backup for when ChatGPT is unavailable." Don't expect it to replace Claude Opus for deep reasoning, but it's more than enough for daily text tasks.

Is it exciting?

What's the 'Wow' factor?:

• True Zero Cost: Free, no ads, no IAP, no subs—almost unheard of for AI apps in 2026.
• Runs in Airplane Mode: The security of knowing it works without internet is something cloud AI can't provide.
• Siri Integration: Just say "Hey, Locally AI" to start chatting without opening the app.

The "Aha!" Moment:

"The new Qwen 3.5 running on-device on iPhone 17 Pro. Qwen 3.5 beats models 4 times its size, has strong visual understanding, and can toggle reasoning on or off." — @adrgrondin (6,906 likes, 3.43M views)

"An LLM is running inside your iPhone, without internet, without API keys, without cloud servers, and it's beating models 4 times its size... In 2025 everyone was arguing about which AI to use; in 2026 the model runs on your phone. The game has completely changed." — @sakevoid (688 likes)

Real User Feedback:

Positive: "For what we are getting — for free no less — far exceeds my expectations" — App Store User Positive: "Truly wonderful... it does not collect any of my data" — InfoSec Analyst Critique: "Voice mode crashes when used on iPhone Air" — App Store User Critique: "Missing DeepSeek V3, GPT OSS, Llama 4 lightweight versions" — App Store User

For Independent Developers

Tech Stack

• Frontend: SwiftUI (Native Apple multi-platform)
• Backend: None, 100% on-device execution
• AI/Inference Engine: Apple MLX Framework (leveraging Apple Silicon unified memory)
• Model Format: MLX format (4-bit/6-bit quantization), sourced from Hugging Face MLX-community
• Voice: 100% local TTS/STT, no cloud processing
• Infrastructure: No servers, zero maintenance cost

Core Implementation

The technical core of Locally AI is Apple's MLX framework. Designed specifically for Apple Silicon, MLX utilizes unified memory (CPU and GPU sharing the same RAM), avoiding the overhead of CPU-GPU data transfer found in traditional frameworks. Once loaded, inference happens directly on the device GPU with minimal latency.

Community tests show Qwen 3.5-2B on iPhone 17 Pro (6-bit quantization) generates roughly 30-50 tokens/sec—comparable to cloud API speeds but without network lag. The "reasoning toggle" is also handled locally, allowing users to switch between fast replies (low latency) and reasoning mode (deep thinking) to manage battery and compute.

Open Source Status

• Is it open source?: No, there is no public repository on GitHub.
• Similar open-source projects: mlx-swift-chat (SwiftUI + MLX chat app), LocalLLMClient (Swift local LLM client library).
• Difficulty to replicate: Low-Medium. The core inference logic is encapsulated by the MLX framework. A basic version could be built in 1-2 person-months. The challenge lies in UX polish (model management, Siri/Shortcuts integration) and continuous adaptation to new models.

Business Model

• Monetization: Currently zero (completely free, no ads, no IAP).
• Potential paths: Pro subscriptions (advanced features), Enterprise versions, or acquisition.
• User base: Not disclosed, but the founder's 3.43M view tweet suggests high brand awareness.

Giant Risk

Medium. Apple built a Foundation Model (LFM) into iOS 26, but its functionality is currently limited to basic chat. Locally AI's advantage is supporting dozens of open-source models with more flexibility and faster updates. The real risk is Apple continuously enhancing native AI to squeeze out third-party apps—though this cycle may take 2-3 years. In the short term, Locally AI's position as a "model supermarket" remains solid.

For Product Managers

Pain Point Analysis

• Problem solved: Lowers the barrier for average users to run local LLMs on mobile (previously a geek-only operation).
• Severity: Medium-High. Privacy anxiety is real (especially under GDPR), and the demand for offline AI grows as model capabilities improve.

User Persona

• Core: Privacy-first users (security pros, lawyers, doctors handling sensitive info).
• Growth: AI enthusiasts/tech early adopters who find "running a model on a phone" inherently cool.
• Potential: Users who want AI but refuse to pay ChatGPT's monthly fee.

Feature Breakdown

Competitor Comparison

Key Takeaways

1. Free strategy builds reputation: Being completely free is a major differentiator in the AI space, quickly accumulating a user base.
2. Siri + Shortcuts integration: Turning local AI into a system-level capability rather than a standalone app significantly increases usage frequency.
3. Compatibility info: Informing users within the app whether a model will run smoothly on their specific device manages expectations.
4. Speed to market: Having an app demo ready the same day Qwen 3.5 launched captured a massive viral window.

For Tech Bloggers

Founder Story

• Founder: Adrien Grondin, Paris, France.
• Background: Full-time iOS developer at MatchGroup (Meetic/Match.com). Previously worked at Stootie and Netatmo. Co-founded travel startup Whimtrip.
• Motivation: Self-proclaimed AI & ML enthusiast; this is a side project. He also built Locally Translate and an MLX voice library.
• Fun Fact: Writes code for dating apps by day, builds AI apps by night. Gave a lightning talk on MLX + Local LLMs at Swift Craft UK. A true one-man army.

Discussion Angles

• "Is mobile AI just a gimmick?": In 2025, articles claimed local LLMs were gimmicks. By 2026, Qwen 3.5-2B running at 30-50 tok/s proves otherwise.
• Model Hallucination: Some users found Qwen 3 claiming to be "GPT-3.5 Turbo by OpenAI and Alibaba." This identity confusion is common in small models but can trigger trust issues.
• Will Apple kill this niche?: iOS 26 has a built-in Foundation Model. Apple has a history of "Sherlocking" third-party apps.

Hype Data

• PH Ranking: 12 votes (Low, likely due to minimal promotion).
• Twitter Viral Success: The founder's Qwen 3.5 demo tweet — 6,906 likes, 652 retweets, 330 replies, 3.43 million views. This is explosive for an independent dev.
• Media Coverage: Recommended by Simon Willison, interviewed on French tech podcasts, featured in MacStories App Debuts.
• Search Trends: "Qwen iPhone local" searches spiked following the Qwen 3.5 release.

Content Suggestions

• The Story: "The French dev building dating apps by day whose free AI app got 3.43M views on Twitter."
• The Trend: "Mobile AI isn't the future—it's happening now (Qwen 3.5 Small series)."
• The Comparison: A video showdown: Locally AI vs. Private LLM vs. Enclave AI.

For Early Adopters

Pricing Analysis

No paid version, no gated features. In an era where every AI app is a subscription trap, this is refreshing.

Getting Started

• Setup Time: 3 minutes.
• Learning Curve: Very low.
• Steps:
  1. Search "Locally AI" on the App Store and download.
  2. Open the app and browse the model list (it highlights which ones fit your device).
  3. Download a model (Qwen 3.5-2B is recommended, ~1.5GB).
  4. Start chatting immediately—no registration required.
  5. Pro tip: Set up Siri integration and create Shortcuts for automation.

Pitfalls & Critiques

1. Voice mode instability: Crashes on iPhone Air and low volume issues. Wait for a patch.
2. Capability Ceiling: Don't expect a 2B model to match GPT-4o. It's great for text editing and translation, but not complex reasoning.
3. Identity Crisis: Qwen 3 might claim to be GPT-3.5 Turbo due to training data quirks.
4. No Document Upload: Unlike Enclave AI, you can't import PDFs/Word docs into the chat yet.
5. Battery Drain: Continuous inference for 2-3 hours will drain an iPhone 17 Pro; intermittent use takes about 15-20% over 8 hours.

Security & Privacy

• Data Storage: 100% local, no internet, no collection.
• Privacy Policy: Developer declares no data collection (confirmed by App Store Privacy labels).
• Audit: No third-party audit, but the offline nature minimizes the attack surface.

Alternatives

For Investors

Market Analysis

• Edge AI Market Size: ~$30-48B by 2026 (Grand View Research).
• Growth Rate: 21-36% CAGR.
• Inference Market: By 2026, AI inference will account for 2/3 of total compute; inference chip market >$50B (Deloitte).
• Drivers: Stricter privacy laws (GDPR), offline needs due to spotty 5G, 30-40% annual gains in mobile chip power, and small models reaching the level of 2-year-old large models.

Competitive Landscape

Timing Analysis

• Why now: Qwen 3.5 Small (March 2026) marks the shift from mobile AI being a demo to being daily-usable. 0.8B models can do OCR; 2B models handle fluent chat; 9B models beat GPT-oss-120B in reasoning.
• Tech Maturity: MLX framework is stable; iPhone 15 Pro+ can run 2B models smoothly. M5 chip Neural Accelerators provide 4x acceleration.
• Market Readiness: High. ChatGPT trained the habit; now users are asking "where is my data?" Local AI is the logical next step.

Team & Funding

• Founder: Adrien Grondin (Solo developer).
• Funding: Likely bootstrapped side project.
• Investment Perspective: Brilliant as an independent project (viral growth, cutting-edge tech), but the solo team and zero revenue are risks. Best viewed as a benchmark for the "Edge AI App" trend rather than a direct investment target.

Conclusion

Bottom line: One of the best free entry points for running local AI on your phone. No cost, no account, no internet—it's worth the download just to have a private backup AI.

Resource Links

2026-03-05 | Trend-Tracker v7.3