Qwen vs DeepSeek: which to run on your iPhone or Mac?

On an iPhone or Mac, Qwen is the more versatile choice and DeepSeek is the specialist. The DeepSeek model a phone can run, DeepSeek R1 (1.5B), is actually built on a Qwen model: DeepSeek trained Qwen 2.5 Math 1.5B to reason step by step like its large R1 model. It’s good at math and logic and weaker at everyday chat. Qwen 3.5, the newest Qwen family for phones, handles everyday chat, writing, many languages and photos, and can also reason when you ask it to.

This comparison is about models you run locally. Both companies also run much larger models in their own cloud apps; those are a different comparison. Sizes and details below come from the model catalog in Modelbase and from the model makers’ pages.

Who makes them

• Qwen is the family of open models from Alibaba’s Qwen team. It spans tiny models for phones up to very large ones.
• DeepSeek is the AI company behind DeepSeek R1, released in January 2025. The full R1 has 671 billion parameters (37 billion active per token) and needs servers. Alongside it, DeepSeek released smaller “distilled” models built on Qwen and Llama (model card). The smallest, DeepSeek-R1-Distill-Qwen-1.5B, is the one that fits a phone.

The models you can run on a phone

Model	Download	Thinking	Photos	Context window	Recommended for
Qwen 3.5 (0.8B)	0.65 GB	Optional	Yes	262,144	iPhone 15 and newer
Qwen 3.5 (2B)	1.75 GB	Optional	Yes	262,144	iPhone 15 Pro and newer
Qwen 3.5 (4B)	3.06 GB	Optional	Yes	262,144	iPhone 15 Pro and newer
Qwen 3 (0.6B)	0.35 GB	Optional	No	40,960	iPhone 14 and older
Qwen 3 (1.7B)	0.98 GB	Optional	No	40,960	iPhone 15 and newer
Qwen 3 Instruct 2507 (4B)	2.28 GB	No	No	262,144	iPhone 15 Pro and newer
Qwen 3 Thinking (4B)	2.28 GB	Always	No	262,144	iPhone 15 Pro and newer
Qwen 3 VL (2B)	1.80 GB	No	Yes	262,144	iPhone 15 Pro and newer
DeepSeek R1 (1.5B)	1.01 GB	Always	No	131,072	iPhone 15 and newer

All are 4-bit MLX models from the Modelbase catalog; “Recommended for” is what the app shows on each model’s card. The context window is the model’s maximum; the app uses as much of it as your device’s memory allows.

Head to head

Everyday chat and writing: Qwen

DeepSeek R1 (1.5B) always reasons before it answers and was built from a math model, so it’s slow and awkward for emails, summaries or small talk. Modelbase’s own description says it’s “weaker at everyday chat.” Qwen 3.5 answers directly when you choose Instant, and is built for general conversation.

Math and logic: both, with different trade-offs

This is DeepSeek R1’s home ground: word problems, arithmetic, algebra, logic puzzles. It shows its reasoning step by step, which you can read. Qwen can do the same: Qwen 3.5 thinks when you set the thinking level to Medium or High, and Qwen 3 Thinking (4B) always reasons. On a phone with enough memory, Qwen 3 Thinking (4B) is a larger reasoning model than DeepSeek R1 (1.5B); on a phone with less, DeepSeek R1’s 1.01 GB is easier to fit.

Photos: Qwen

Every Qwen 3.5 model can look at images, so you can ask about a photo, a menu or a screenshot (how that works). DeepSeek R1 (1.5B) is text only.

Languages: Qwen

Modelbase’s catalog describes Qwen 3.5 (4B) as writing in over 200 languages. DeepSeek R1 (1.5B) was built for math reasoning rather than languages, so for chatting in anything other than English, Qwen 3.5 is the safer choice.

Long conversations and documents: Qwen 3.5

Qwen 3.5 and Qwen 3 (2507) models accept up to 262,144 tokens, DeepSeek R1 (1.5B) up to 131,072. In practice your device’s memory sets the limit for both, but Qwen leaves more room.

License: DeepSeek is MIT

DeepSeek’s distilled models are released under the MIT license. Qwen models come with their own licenses; check the model page if you plan to build on one.

Time and battery: Qwen with Instant is lighter

A reasoning model writes out its thinking before the answer, and every word of that thinking is work for your phone’s chip. DeepSeek R1 (1.5B) always does it, so even a simple question takes longer and uses more battery than the same question to Qwen 3.5 with the thinking level on Instant. That’s the right trade for a tricky word problem and the wrong one for “rewrite this message more politely”. With Qwen 3.5 you choose per question: Instant for everyday things, Medium or High when a problem deserves it. Modelbase keeps the reasoning in a line above the answer that you can open, so you can see how either model got there, which is useful when you’re checking math. (More in what a reasoning model is.)

Which should you pick?

You want	Pick
One model for everything on a recent iPhone	Qwen 3.5 (2B)
The most capable Qwen your phone can hold	Qwen 3.5 (4B)
Something light for an older iPhone	Qwen 3 (0.6B)
Step-by-step math and logic, small download	DeepSeek R1 (1.5B)
Stronger step-by-step reasoning, more memory	Qwen 3 Thinking (4B)
Questions about photos	Any Qwen 3.5

You can also keep two: a Qwen for everyday use and DeepSeek R1 for math. Models are a download away, and you can switch between them in any chat.

How to try both in Modelbase

1. Install Modelbase (free, iPhone, iPad and Mac; no account).
2. Tap the model’s name at the top of a chat, then Manage models > Local.
3. Search for “Qwen” or “DeepSeek” and tap Download on the models you want.
4. Pick one in the model picker. For reasoning, set the thinking level next to the message field to Medium or High.

More on running DeepSeek in How to run DeepSeek R1 locally on iPhone and Mac, and on choosing between all the families in Gemma vs Llama vs Qwen.

FAQ

Is DeepSeek based on Qwen?

The small DeepSeek models are. DeepSeek-R1-Distill-Qwen-1.5B and 7B are built on Qwen 2.5 Math, and the 14B and 32B on Qwen 2.5; the 8B and 70B are built on Llama. The full DeepSeek R1 is DeepSeek’s own model.

Which is better, Qwen or DeepSeek?

On a phone, Qwen 3.5 is better for most things (chat, writing, photos, languages); DeepSeek R1 (1.5B) is a focused choice for step-by-step math and logic. For the large cloud models, compare them on the providers’ own sites.

Can I run DeepSeek and Qwen offline?

Yes. Once downloaded in an app like Modelbase, both run entirely on your iPhone, iPad or Mac, with no connection.

Which uses less memory?

Qwen 3 (0.6B) at 0.35 GB is the smallest; DeepSeek R1 (1.5B) is 1.01 GB; Qwen 3.5 (2B) is 1.75 GB. A model needs about its download size, and some more, in free memory while it answers.