Qwen Chat from Alibaba Now Runs Fully Without Internet — Even on Budget Smartphones

Alibaba's newly released Qwen 3.5 Small Series runs entirely on-device, requiring no Wi-Fi, no account, and no subscription after a one-time download, a landmark shift in how AI assistants work on consumer hardware.

The biggest limitation of modern AI tools like ChatGPT, Gemini, or Claude is simple: they require the internet. Every prompt you type travels to remote servers, where the model processes it and sends back the response.

But a new development from Alibaba is challenging that assumption.

With the release of the Qwen 3.5 model family and its lightweight variants, developers have started running Qwen Chat completely offline on smartphones and laptops. Once downloaded, the AI continues to function even when the device is in airplane mode, meaning no data leaves the device and no internet connection is required.

This shift represents a major milestone in the AI industry: powerful generative AI that runs locally on your own device.

Qwen Chat: At a Glance

• Released March 2, 2026 — four models ranging from 0.8B to 9B parameters
• Licensed under Apache 2.0 — free for personal and commercial use
• The 2B model runs on phones with 6GB RAM (~$200–300 Android range)
• Supports 201 languages and dialects natively
• 262,144-token context window — far exceeding most cloud models' free tiers
• 9B model outperforms OpenAI's gpt-oss-120B on key benchmarks despite being 13× smaller

What Qwen Chat Is - And What Changed

Qwen Chat is Alibaba's flagship conversational AI interface, available at chat.qwen.ai for cloud-based use. Like ChatGPT or Claude, it offers document understanding, image analysis, web search, and code generation. But unlike those services, the underlying model weights have now been released publicly — small enough, for the first time, to run on the device already in your pocket.

The 2B parameter variant is the key unlock. At its most compressed quantization, it requires as little as 528MB of storage. That file size, combined with a minimum RAM requirement of 6GB, means the model runs on mid-range Android phones costing between $200 and $300. iPhone users can access it through apps like PocketPal AI and Off Grid, which support Apple Neural Engine acceleration on A-series chips.

The Four Models: A Breakdown

The Qwen 3.5 Small Series ships in four configurations, each targeting a different tier of consumer hardware.

The 9B model is the standout performer. Despite having just 9 billion parameters, it surpasses OpenAI's open-source gpt-oss-120B — a model 13 times its size — on several independent benchmarks, including graduate-level science reasoning (GPQA Diamond: 81.7) and visual understanding (Video-MME with subtitles: 84.5). On Apple Silicon iPhones, it runs at approximately 30–45 tokens per second — fast enough to feel conversational.

Performance: How Fast is Offline Qwen AI?

Running AI locally is impressive, but there are trade-offs.

On a typical smartphone:

• The 2B model runs at around 8 tokens per second.

This is slower than cloud AI models but still usable for many tasks.

Typical response times:

Performance improves significantly on:

• Modern smartphones
• GPUs
• Apple Neural Engine devices

The Architecture Behind the Efficiency

The performance-to-size ratio isn't accidental. Alibaba's researchers replaced approximately 75% of the standard Transformer attention layers with a mechanism called Gated DeltaNet — a linear attention architecture that scales with sequence length far more efficiently than traditional quadratic attention. This is how Qwen 3.5 achieves a native 262,144-token context window without the memory cost that would make that impractical on a phone.

The "gated" element combines two memory-control mechanisms: a decay gate for selectively forgetting accumulated state, and a delta rule for surgical key-value updates. Together, they solve the quality degradation that plagued earlier linear attention designs, allowing long documents to be processed accurately at a fraction of the compute cost.

All four models were also trained using early fusion — text, images, and video present together from the start of training, not bolted on afterward. This is why even the 4B model handles visual content at a level previously requiring models many times its size.

How to Set It Up: Step-by-Step

There are two primary apps for running Qwen 3.5 offline: PocketPal AI (fully open-source, available on both iOS and Android) and Off Grid (broader feature set, supports voice, image generation, and tool calling). The setup process is identical for both.

1. Download the app: PocketPal AI or Off Grid from the App Store (iOS) or Google Play (Android). Both are free.
2. Connect to Wi-Fi: The model files range from 528MB to several GB. Download over Wi-Fi to avoid mobile data charges.
3. Select your model: For most phones with 6GB RAM, choose Qwen 3.5 — 2B. For 8GB+ or Apple Silicon, try the 4B. The 9B is for high-end devices only.
4. Wait for the download to complete: Depending on connection speed and model size, this takes 1–10 minutes.
5. Enable airplane mode to verify: Turn off all network connectivity. Open the app, ask a question. It responds — nothing was sent anywhere.
6. Configure to your use case: Adjust temperature, context length, and system prompts in the app settings. PocketPal allows custom AI personas for different tasks.

Why Offline AI Is a Bigger Story Than It Sounds

The privacy case for on-device AI is straightforward but worth stating plainly: when you use ChatGPT, Claude, or Gemini, your messages travel to a remote server. That transmission creates a log. With offline Qwen, there is no transmission. Conversations about sensitive topics — health conditions, legal questions, business strategy, personal relationships — stay on the device permanently.

The connectivity argument is equally concrete. A working AI assistant on a flight with no Wi-Fi. In a hospital where cloud service compliance is fraught. In regions where AI platform APIs are geographically restricted. Underground. In rural areas with no signal. Once downloaded, the model requires none of these conditions.

And the economics matter too. There is no monthly subscription. No per-token billing. No free tier that expires. The Apache 2.0 license permits free personal and commercial use. For students, developers in cost-sensitive markets, and high-volume users, this is a material difference from every major cloud AI service.

How Qwen 3.5 Compares to Other Offline Options

Latest AI NEWS