Alibaba launches Qwen3.8-Max with 2.4T parameters and 1M-token context window
Alibaba has launched Qwen3.8-Max, its largest AI model to date, through QwenCloud. The mixture of experts model has 2.4 trillion total parameters but activates 95 billion per request to reduce inference costs and response times. It supports text, images, and video, with a context window of up to 1 million tokens for processing large documents, codebases, and extended task histories.
Alibaba is positioning Qwen3.8-Max primarily for coding and cowork tasks that require AI agents to operate independently for extended periods. The company says the model completed an internal software project over 16 days and can continuously handle autonomous development work for more than 10 days. Alibaba also claims it outperformed OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5 across several coding, agentic, general, and multimodal benchmarks, although the results have not been independently verified.
The model is currently available through QwenCloud, while its weights are scheduled for release on Hugging Face and ModelScope next week. The launch follows other large Chinese AI models like Kimi K3 from Moonshot AI as competition with US laboratories continues to increase.
- AI Chatbot
- Free
- Open Source
Related news
External links
- Qwen3.8-Max: A New Bar for Coding and CoworkQwen Blog • Official source
- Alibaba unveils Qwen 3.8 Max as DeepSeek undercuts AI rivals with ultra-low-cost modelFirstpost
- Alibaba releases Qwen3.8-Max, challenging GPT-5.6 Sol and Claude Fable 5 on AI benchmarksNeowin
- Alibaba Releases Qwen 3.8 Max, Beats GPT 5.6 Sol And Fable On Many BenchmarksOfficeChai
Comments
The article's own caveat matters most here — "outperformed X and Y" with no independent verification is doing a lot of work in that sentence. Self-reported benchmark comparisons from the company launching the model are worth treating as a marketing claim until a third party reproduces them, not as an established ranking.
The 95B active out of 2.4T total is the more interesting engineering number than the headline parameter count, honestly, that ratio is what actually determines inference cost and latency in practice, and it's easy for "2.4 trillion parameters" to imply a cost profile the model doesn't actually have once you know only ~4% activates per request. The 16-day autonomous run is the claim I'd want the most detail on before treating as a capability signal rather than a demo, a long uninterrupted run is compatible with either genuine sustained productive work or a well-orchestrated loop that looks impressive without much human-reviewable substance in between. Worth the same skepticism the piece already applies to the benchmark numbers.
This is a significant release from Alibaba. A 2.4-trillion-parameter model with a 1 million-token context window shows how quickly frontier AI is advancing. It will be interesting to see how Qwen3.8-Max performs in real-world coding, reasoning, and multimodal tasks compared with leading models from OpenAI, Anthropic, and other competitors, especially once more independent benchmarks and developer feedback become available.
"the open weights will be released next week." It's not the first time Alibaba announces opening models then never has (and finally announcing closing them), but since Kimi K3 is now also open weight and is very popular (to the point that OpenAI is discounting GPT-5.6), the room for closed models is shrinking every day.
It's time to rethink the whole AI economics, and banning Chinese models in the US (even if it's technically impossible) would only hurt OpenAI and Anthropic even more.