Alibaba releases open Qwen3.8 with 2.4T total, 95B active parameters

Alibaba released the open-source Qwen3.8-2.4T-A95B model weights on August 12, a 2.4-trillion-parameter sparse mixture-of-experts model with about 95 billion active parameters per token, making it the first Qwen-Max-class model released openly. The text-only checkpoint requires thinking mode, supports a 262,144-token native context extendable to 1,010,000 tokens, and is compatible with Hugging Face Transformers, vLLM, SGLang, and TokenSpeed. The hosted Qwen3.8-Max on Qwen Cloud adds vision input, non-thinking mode, and built-in tools, but independent benchmarks for the open checkpoint are not provided.

Developers and infrastructure teams can now download the model weights for Qwen3.8, the latest release from Alibaba https://home.alibabagroup.com/en-US/ir-corporate-governance-partnership?ref=runtimewire 's large-model program. Alibaba published the Qwen3.8-2.4T-A95B weights https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B?ref=runtimewire on August 12. Qwen calls it the first Qwen-Max-class model released openly in its Qwen3.8 announcement https://qwen.ai/blog?id=qwen3.8&ref=runtimewire . The repository provides Qwen3.8-2.4T-A95B as model weights and configuration files in Hugging Face Transformers format, with compatibility for vLLM, SGLang and TokenSpeed https://lightseek.org/tokenspeed/recipes/models?ref=runtimewire qwen3-8 . Qwen separately offers Qwen3.8-Max through Qwen Cloud https://www.qwencloud.com/models/qwen3.8-max?ref=runtimewire . According to Qwen's repository and cloud documentation, that hosted version supports vision input, non-thinking mode and built-in tools. What Qwen released Qwen's repository describes Qwen3.8-2.4T-A95B as a 2.4-trillion-parameter sparse mixture-of-experts model with about 95 billion parameters activated per token. Its 92 layers contain 512 experts, with 10 routed experts and one shared expert active at a time. That design increases the model's stored capacity without using all 2.4 trillion parameters for every inference step. The repository says the open checkpoint is text-only, requires thinking mode and supports a native 262,144-token context that can be extended to 1,010,000 tokens. Its reasoning effort setting supports xhigh by default, medium and low . A separate preserve thinking control retains reasoning context from earlier messages. The model card lists support for Hugging Face Transformers, SGLang https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8?ref=runtimewire , vLLM https://recipes.vllm.ai/Qwen/Qwen3.8-2.4T-A95B?ref=runtimewire and TokenSpeed. This is far removed from the local-development case for smaller Qwen releases. RuntimeWire reported in June that Qwen 3.6 27B had become practical for local coding work /article/qwen-36-27b-local-development-piotr-migdal-quesma . Qwen3.8's large parameter count creates a substantial storage requirement before an operator accounts for serving memory, compute and inference throughput. The open checkpoint is narrower than Qwen3.8-Max The downloadable checkpoint is text-only and always operates in thinking mode. It does not accept image or video inputs, and developers cannot disable its reasoning mode through the documented interface. According to Qwen's model repository https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B?ref=runtimewire , Qwen3.8-Max adds vision input, non-thinking mode, built-in tools and a default 1-million-token context window. The same repository directs users seeking managed, scalable inference without infrastructure maintenance to Qwen Cloud. The two releases give developers a downloadable text model and a hosted Max product with additional input modes and tools. The repository links to a model-specific license The repository links to a Qwen3.8-specific license file https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE?ref=runtimewire . The supplied materials do not include the license text, so its commercial-use conditions cannot be summarized here. What the published benchmarks establish The supplied materials do not establish independent benchmark results for the downloadable Qwen3.8-2.4T-A95B checkpoint, while Qwen's model card reports company-run comparisons for the hosted Qwen3.8-Max version https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B?ref=runtimewire . Those comparisons cover coding, agent and general-capability tests under varying harnesses, run counts and evaluation conditions. Performance and operating costs for third-party deployments of the downloadable checkpoint therefore remain unestablished by the supplied evidence.