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

Alibaba has launched the open weights for Qwen3.8-27B and Qwen3.8-2.4T-A95B, marking the first time the company has made a flagship Max-class model downloadable. Released under the Apache 2.0 license on Hugging Face and ModelScope, these models target different hardware budgets. The Qwen3.8-2.4T-A95B is a massive mixture-of-experts model featuring 2.4 trillion total parameters with 95 billion active parameters per token, keeping inference costs manageable. Meanwhile, Qwen3.8-27B is a compact, deployment-friendly dense model designed for local VRAM predictability and low latency.

To support a native 262K-token context window without quadratic memory costs, the Qwen3.8-27B model utilizes a hybrid Gated DeltaNet architecture. It features 64 layers organized into 16 repeating blocks. In each block, three Gated DeltaNet sublayers use linear O(n) attention, while only one sublayer uses full Gated Attention to maintain long-range precision. This design allows the dense model to handle massive context windows efficiently.

The Qwen3.8-27B model outperforms the API-only Qwen3.7-Plus on key benchmarks. It scores 61.7 on SWE-bench Pro for agentic coding compared to 57.6 for Qwen3.7-Plus, and 70.7 on CoWorkBench for office tasks compared to 65.1. Both new models feature a thinking mode enabled by default, offering tunable reasoning effort levels of xhigh, medium, and low. They also include a preserve_thinking flag to maintain coherence across multi-turn agent interactions.

For practitioners, this release provides unprecedented flexibility to self-host a highly capable 2.4-trillion-parameter model or run a highly efficient 27B model locally. For those preferring cloud access, the Max model API is priced on QwenCloud at $2 per million input tokens and $6 per million output tokens, with a hosted API for the 27B model coming soon.