開放模型與推論基礎設施
Qwen3.8 Makes 2.4T-Parameter Max-Class Weights Available for the First Time, Activating 95B per Token
Alibaba’s Qwen team has released Qwen3.8-2.4T-A95B, featuring 512 experts and a hybrid of linear and standard attention, with support for contexts of up to approximately 1.01 million tokens. The model offers adjustable reasoning depth, but accepts text only and cannot disable thinking mode, making its practical deployment requirements far higher than those of typical open-weight models.
Alibaba’s Qwen team released Qwen3.8-2.4T-A95B on August 12, marking the first time it has made the weights of a Max-class model available to the community. The model has 2.4 trillion parameters in total and activates 95B parameters per token. Its 92-layer network is arranged into 23 blocks, each containing three layers of Gated DeltaNet linear attention and one layer of standard Gated Attention. The MoE layers contain 512 experts, selecting ten routed experts and one shared expert for each token, and the model is trained with multi-token prediction. Its native context length is 262,144 tokens and can be extended to approximately 1.01 million tokens.
The open-weight version is currently text-only, and every response must use thinking mode. `reasoning_effort` offers three levels—low, medium, and xhigh—while `preserve_thinking` retains earlier reasoning across multi-turn messages. These interfaces are especially important for long-running agents: engineers can trade reasoning budget for latency and cost without rewriting prompts or switching models. The weights are compatible with vLLM, SGLang, and TokenSpeed, but the 2.4T-parameter BF16 files mean that a full deployment requires multiple nodes, expert parallelism, and high-capacity storage. “Only 95B parameters are activated per token” should not be misinterpreted as meaning that only 95B parameters need to be loaded.
Official evaluations show Qwen3.8-Max scoring 86.6 on Terminal-Bench 2.1 and 67.7 on SWE-bench Pro, up from Qwen3.7-Max’s 74.5 and 60.6, respectively. It also reports a score of 93.0 on PaperBench. However, several results were obtained using Claude Code, OpenCode, or Qwen’s own harnesses, and some benchmarks rely on internal data, so they do not constitute a fully consistent apples-to-apples comparison. The open-weight model is also not identical to the cloud-hosted Qwen3.8-Max, which additionally supports visual input, a non-thinking mode, a default 1M-token context window, and built-in tools. The next things to watch are independent reproductions, throughput under different expert-parallel configurations, and whether agent success rates hold up when the reasoning budget is reduced.