The Alibaba Qwen team has published the open weights for the Qwen3.8-2.4T-A95B model, presenting it as the first Qwen-Max class model available for download, not just via API. This sparse mixture-of-experts checkpoint contains approximately 2.4 trillion shared parameters and activates about 95 billion parameters per token across 512 experts, utilizing 10 routed experts plus one shared expert at each step.
The release of this model differs from previous open Qwen releases, such as Qwen-Drive and smaller Qwen3.8 code snapshots. It positions the large-scale text MoE for use in proprietary environments for agentic tasks and reasoning, rather than as a specialized perception or encoding model.
The model's architecture is based on a hybrid attention system that alternates between gated linear layers and full attention over 92 layers. Most layers utilize limited recurrent state instead of a growing key-value cache, while a small portion retains full pairwise attention for high-precision interactions. This approach aims to keep computational costs and memory consumption manageable as the context window scales from a native size of around 262 thousand tokens to approximately one million with expansion. The maximum output sequence length is up to 128 thousand tokens.
Built-in reasoning control mechanisms allow developers to regulate the effort expended per query, trading latency for deeper multi-step traces in coding, research, long document analysis, and tool-using workflows that accumulate system prompts, tool outputs, and intermediate plans.
Deployment services were quickly released alongside the weights. Amazon Web Services published a SageMaker HyperPod guide demonstrating model deployment using vLLM on ml.p6-b300 instances. This guide covers cluster setup, NVFP4 quantization enabling weight placement on an eight-GPU node, tool calling, and speculative decoding with multi-token prediction running behind an OpenAI-compatible endpoint.
NVIDIA also provided recipes for SGLang, vLLM, and Dynamo on GB300 NVL72 hardware, reporting speeds exceeding 4000 tokens per second per GPU and over 350 tokens per second per user when using FP8 format in internal tests. NeMo AutoModel paths are available for subsequent supervised fine-tuning or LoRA on Hugging Face checkpoints. Furthermore, some inference providers offer hosting options.
The open checkpoint file is available only in text format under the Qwen3.8-Max license with commercial terms for large model service operators. However, the cloud-provided Qwen3.8-Max product may differ in visual and modal configurations. According to vendor benchmarks, the model shows strengths in research and instruction following, while potential for improvement is noted in more complex repository tasks.
For teams requiring Max-class power within their own infrastructure, the practical takeaway is the combination of downloadable weights and ready-made recipes for both cloud and open hosting, available concurrently. Thus, the 2.4T MoE can now be run where operators already use standardized solutions like vLLM and managed GPU clusters.