Alibaba outsmarts DeepSeek in AI offerings
I might be wrong, but I believe that Alibaba’s Qwen offerings have become more interesting than DeepSeek’s. At least, when coding is what matters the most to you.
When I first mentioned Qwen3, I noted that the Android app was not available “for my device” but most likely not available outside a few Asian countries. A QR code allowed for the download of version 1.0.0 of the APK, but that was frustrating.
The other day, I noticed that the default model in the web app became Qwen3-Coder, and the Android app asked me to download an update. There is a new page for those wishing to Download Qwen, with links for Android and macOS (no iOS). The app in Google Play Store is still “not available for any of [my] devices”; however, the QwenChat_release.apk offered by the above page is now at version 1.4.0, which is currently the latest.
The default model is now Qwen3-Coder:
- Maximum context length: 1,048,576 tokens
- Maximum generation length: 65,536 tokens
The flagship model has been updated to July, Qwen3-235B-A22B-2507:
- Maximum context length: 131,072 tokens
- Maximum summary generation length: 8,192 tokens
- Maximum generation length: 81,920 tokens
Obviously, Qwen is betting on code generation.
However, the Building AI products with Qwen page seems out-of-sync. The shown context lengths don’t match the above models. The Alibaba Cloud Model Studio is where the truth can be found. 😉
The featured models are the newest ones.
qwen3-coder-plus
Pricing here.
- Context window 1,048,576
- Maximum input (tokens) 1,000,000
- Maximum output (tokens) 65,536
Tiered pricing:
Input token count Input price (Million tokens) Output price (Million tokens) 0-32K $1 $5 32K-128K $1.8 $9 128K-256K $3 $15 256K-1M $6 $60 qwen-plus-2025-07-14
Price per 1M tokens:
- Input $0.4
- Output $1.2
- Output (Thinking) $4
qwen-omni-turbo
Pricing here.
Qwen-Omni supports multiple input modalities, including video, audio, image, and text. It can output audio and text.
Price per 1M tokens when the input and output are text-only:
- Input $0.07
- Output $0.27
Note that the coder isn’t that cheap, but with real code, not just toy projects, the huge context is important.
Many other models are available, and you can even compare them 3 at a time. Example:
These “plus” models aren’t available in the web app or mobile app, but only via API calls. Fair enough. The 1 million token context Qwen3-Coder can nonetheless be tested via the browser.
🐼
If I want to compare Alibaba’s with DeepSeek’s offerings, DeepSeek’s website falls short of convincing: there’s a link to the web app (which is not bad, thank you), another link to the mobile app (it’s not a true link as it only shows a QR code which goes to download.deepseek.com/app/), and a link to their API platform.
The Models & Pricing page kind of disappoints. Still, it’s nice that they have this offer:
Off-Peak Discounts：DeepSeek-V3 with 50% off and DeepSeek-R1 with 75% off at off-peak hours (16:30-00:30 UTC daily).
- The
deepseek-chatmodel points to DeepSeek-V3-0324. - The
deepseek-reasonermodel points to DeepSeek-R1-0528. - The maximum input context length is 64K for both models, which means they cannot compete with Alibaba’s Qwen models! Also, to cut costs, you should use context caching.
- The maximum output context length is 8K for
deepseek-chatand 64K fordeepseek-reasoner(including the Chain-Of-Thought!), with the default values being half of the maximum (this is what I’d expect in the web app and mobile app).
The standard (not discounted) prices per 1M tokens:
| Model | deepseek-chat | deepseek-reasoner |
| Input (cache hit) | $0.07 | $0.14 |
| Input (cache miss) | $0.27 | $0.55 |
| Output | $1.10 | $2.19 |
🐼
I wonder how many people outside China (or outside BRICS) used a paid account with either of Alibaba or DeepSeek to get code assistance via API. Do they trust the Chinese?
Not that Anthropic, OpenAI, or xAI should be trusted with your data and your code… (Or Amazon, if you’re using Claude via Kiro.)
Either way, at the annual World Artificial Intelligence Conference in Shanghai, China just proposed a new global AI cooperation organization.
🐼
UPDATE: Oh, but there is more! Kimi and Z.AI: The more Chinese, the merrier!
The update to Qwen3-235B-A22B-2507 is seen by The Reg as follows: Alibaba admits Qwen3’s hybrid-thinking mode was dumb.
Of course, Alibaba admitted as much:
Back to The Reg:
I only used Qwen in a browser, in some weeks more than in the 2nd week of September, when it only accounted for 9% of the interactions. Lately, tough questions went to Copilot and ChatGPT, other questions to Claude and Grok, and web searches to Qwen and Kimi.
But on Android I was still having version 1.4.0 of the app, because it’s not available on the Google Play Store, so I needed to update it manually from download.qwen.ai/android/QwenChat_release.apk.
In version 1.8.0, I noticed that there are 20 available models:
But is Qwen3-Max really better than Qwen3-235B-A22B-2507, the model I was used to?
They say Qwen3-Max has over 1 trillion parameters and was pretrained on 36 trillion tokens. So maybe I should believe them when they claim it outperforms Qwen3-235B-A22B-2507, which is a Mixture-of-Experts (MoE) model with 235 billion parameters but only 22 billion active parameters per inference.
But 1 trillion, really?! They cannot all be active for a single task!
I didn’t even notice when the default model changed from Qwen3-Next-80B-A3B to Qwen3-Max! (Prior to that, it was Qwen3-235B-A22B-2507.)