started · updated
Alibaba launches Qwen4-Max AI model with 2.4 trillion parameters
Alibaba has released Qwen4-Max, its latest flagship large language model, designed to handle complex professional and scientific tasks. The model utilizes a Mixture-of-Experts (MoE) architecture, featuring 2.4 trillion total parameters with 71 billion active parameters per token.
Key improvements include a 1-million-token context window and enhanced capabilities in 3D modeling, finance, healthcare, and coding. Alibaba claims the model outperforms its predecessor, Qwen3.8-Max, while significantly reducing token usage and operational costs. In specific evaluations, the average cost per task reportedly fell by approximately 62%.
Security is a central focus of the release. Alibaba stated it filtered training data to remove offensive cyberattack capabilities while retaining data for defensive tasks like vulnerability detection. The model is currently available via API and through Alibaba Cloud, with smaller versions expected to be released as open-weights in the future.
Entities
Alibaba · Alibaba Cloud · Anthropic · China · Claude Opus 5.5 · OpenAI · Qwen · Qwen4-Max · Qwen4-Max · QwenCloud
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 7 SOURCES] Qwen4-Max outperforms Qwen3.8-Max in nearly every evaluated capability. analyticsindiamag.com · www.tekedia.com · pc.watch.impress.co.jp · www.blogspan.net · www.infoq.cn · +2 more
- [● 16 SOURCES] Qwen4-Max has 2.4 trillion total parameters. analyticsindiamag.com · jornaleconomico.sapo.pt · pc.watch.impress.co.jp · www.blogspan.net · the-decoder.de · +11 more
- [● 13 SOURCES] Qwen4-Max uses a Mixture-of-Experts architecture with 71 billion active parameters per token. pc.watch.impress.co.jp · www.blogspan.net · www.frandroid.com · the-decoder.de · zdrojak.cz · +8 more
- [● 6 SOURCES] Alibaba filtered training data to remove offensive cyberattack capabilities while keeping defensive functionalities. analyticsindiamag.com · jornaleconomico.sapo.pt · sapo.pt · srnnews.com · hojemacau.com.mo · +1 more
- [● 6 SOURCES] The model is a closed model available only via API. jornaleconomico.sapo.pt · www.blogspan.net · zdrojak.cz · www.hwupgrade.it · www.frandroid.com · +1 more
- [● 7 SOURCES] The model's international API price is $1.50 per 1M input tokens and $4.50 per 1M output tokens. dev.to · exame.com · pc.watch.impress.co.jp · www.frandroid.com · www.hwupgrade.it · +2 more
- [● 8 SOURCES] Alibaba released the Qwen4-Max model on November 16, 2026. dev.to · pc.watch.impress.co.jp · news.mydrivers.com · cryptobriefing.com · www.infoq.cn · +3 more
- [● 13 SOURCES] The model features a 1-million-token context window. dev.to · analyticsindiamag.com · exame.com · pc.watch.impress.co.jp · www.blogspan.net · +8 more