Alibaba Releases Qwen3.8-Max and Qwen3.8-27B Weights

Alibaba has released weights for Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model activating roughly 4 percent of its parameters, or 95 billion per token. Alongside it, the company released the smaller Qwen3.8-27B. This marks a shift from Alibaba's previous strategy of keeping Max models proprietary while only releasing weights for smaller models like Qwen3.6-27B. While Qwen3.8-Max supports up to a 1-million-token context window, 131,000 output tokens, and 262,000 reasoning tokens at 77.6 tokens per second, its downloadable weights are text-only. These weights are free under a custom license unless a developer exceeds 100 million monthly active users, $20 million in monthly revenue, or $50 million in annual revenue for model-as-a-service platforms. Meanwhile, Qwen3.8-27B is available under a standard Apache 2.0 license.

On the Artificial Analysis Intelligence Index, Qwen3.8-Max scored 58 ($1.13 per task), ranking fifth overall and second among open models. This placed it ahead of Qwen3.7-Max (47, $0.54 per task), Claude Opus 4.8 (57, $2.03 per task), and Muse Spark 1.2 (57, $0.40 per task), but behind Kimi K3 (60, $0.86 per task) and GLM-5.3 (60). On the 𝜏³-Banking test, Qwen3.8-Max scored 51.3 percent. On GDPval-AA v2, it achieved 1,739 Elo, trailing Claude Opus 5 (1,846 Elo) and Claude Fable 5 (1,743 Elo). On Arena.ai's Vision Arena, it scored 1,301 Elo, behind Claude Fable 5 (1,315 Elo). On WebDev Code Arena, it scored 1,667 Elo, behind Claude Opus 5 (1,686 Elo) and Kimi K3 (1,675 Elo). Qwen3.8-27B scored 52 on the Artificial Analysis average, comparable to GPT-5.6 Luna.

For practitioners, Alibaba Cloud Model Studio offers the Qwen3.8-Max API at $2.00 per million input tokens, $0.25 per million cached tokens, and $6.00 per million output tokens. Although hosting a 2.4-trillion-parameter model is impractical for most teams, the release of these weights will drive down costs through third-party hosting competition. Furthermore, developers can run the highly capable Qwen3.8-27B locally on consumer hardware, providing a powerful offline alternative to proprietary systems.