Summary:
- Open weights Qwen-Max-class model for downloads around the world. Here.
- Build on Qwen 3.5.
- 2.4 trillion parameters.
- Available on Qwen Cloud.
- Evaluated on standard LLM datasets, the evaluation is promising when results are compared to Opus4.8, Fable5, GPT5.6 Sol, and Qwen3.7-Max, Qwen’s last model. Not bad I would say. We need cybersecurity tests: CyberLLMInstruct, SecBench, which have not yet been provided as of now while Meta’s CyberSecEval, is available but the model has been evaluated for cybersecurity, especially for the CVE Discovery Benchmark. This model reportedly found more CVEs than most models being tested to date. Still, baselines and gold standards matter in an area like security.
- Someone needs to provide charts for CyberLLMInstruct, and SecBench on your own Qwen instances; please let me know! I don’t want to run a Chinese model, which is new in the market.
- Don’t go by open weights and price alone; consider safety, too. Open weights are there for the Meta Muse Glimmer 30-billion-parameter model. Use that instead.
Qwen is a large language model developed by the Chinese firm Alibaba Cloud. The same firm that owns a wholesale marketplace in China connecting manufacturers and wholesale suppliers. From electronics to home decor, textiles, and everything you can think of in the wholesale market, Alibaba sells that. Its net worth is $296 billion USD. It has money to power the AI Cloud firm, and Qwen is part of it. So no doubt this would be a tough, robust model, since itis built under Alibaba. This model is near the best American models on the market in tasks such as reasoning and coding.
Now Alibaba Cloud has released many LLM models under the Qwen brand. Qwen already has around 3 billion downloads. But don’t go by the numbers yet. China’s population is 1.4 billion, with around 1 billion in the young and working-age group. So if one person downloads two models, this makes a huge difference. The World Artificial Intelligence Cooperation Organization (WAICO) has 29 countries, and they all support China’s AI approach. This can amount to billions of downloads. But do all of these countries know how to use AI? Are these 29 countries AI adults? Does this mean these countries’ population is AI-exposed? We need explict tests on CyberLLMInstruct, SecBench, and Meta’s CyberSecEval, which have not yet been provided. However, CVE Discovery Benchmark, Jailbreak & Robustness Testing have been performed.
Regarding other users of AI for Qwen, I don’t think EU laws should allow Qwen into the market so soon, because EU laws, however strong, haven’t done as extensive cyber testing as it should have. EU regulations are stringent, but do they test it as US governments do? Member countries must run their tests as well. When Mythos 5 and OpenAI can run out of systems without internet, why can’t Chinese Qwen? Who shall be the deployer and who shall be the importer of these models in the EU? Who shall be the notification agent? EU demands a transparency act for now. But Qwen is in the EU. Qwen is accessible in Europe via Amazon Bedrock.
Can we get the demographics of which countries download the most Qwen models?
Alibaba has All in One AI + Cloud, which integrates it all. But to be frank, I don’t want to try a Chinese AI model, since I still need to review charts on Qwen’s AI security. Don’t go by just downloads. Look at the numbers too, especially the error rate. Look at the lab tests: can it leak from lab tests? It’s a big question before starting to use Qwen3.8 Max. Don’t go by open weights and price alone; consider safety as well. Open weights are there for the Meta Muse Glimmer 30-billion-parameter model. Use that instead.
Thank you for reading.
Subscribe for updates.
Artificial Intelligence and Future 