Alibaba Open-Sources Qwen3.8-2.4T-A95B, Its First Qwen-Max-Class Model With Public Weights

Alibaba Releases Its Largest Open-Weight Model Yet

Alibaba's Qwen team released Qwen3.8-2.4T-A95B on August 12, 2026. This marks the first time a Qwen-Max-class model has been made available as open weights, giving developers direct access to a model architecture previously reserved for Alibaba's proprietary API tier.

The model carries 2.4 trillion total parameters but activates only 95 billion per token, a mixture-of-experts (MoE) design intended to keep inference cost tied to activated parameters rather than the full weight count. Weights are published on Hugging Face in standard Transformers format.

Architecture Details

Qwen3.8-2.4T-A95B uses a fine-grained MoE with 512 routed experts plus one shared expert, activating 10 routed experts per token across 92 layers. The layer layout repeats a pattern of three Gated DeltaNet (linear attention) layers followed by one Gated Attention (full quadratic attention) layer — a 3:1 ratio that keeps compute and memory bounded as context length grows. Of the 92 layers, 69 use linear attention with a fixed-size recurrent state instead of a growing KV-cache, while 23 use full attention for higher-fidelity token interactions.

Native context window is 262,144 tokens, extensible to 1,010,000 tokens according to Alibaba. Maximum output length is 128K tokens. The model includes native Multi-Token Prediction (MTP) draft heads, enabling speculative decoding without a separate draft model, and a reasoning_effort parameter (low, medium, high) that lets developers trade inference compute for reasoning depth on a per-request basis.

Benchmark Claims

According to Alibaba's own benchmarking, Qwen3.8-2.4T-A95B scores 93.0 on PaperBench (research workflows), 82.8 on IFBench (instruction following), and 86.6 on terminal-based coding tasks. The company claims the model performs comparably to leading frontier models across most categories, with acknowledged gaps remaining on harder repository-level tasks (SWE-bench Pro) and general tool use (Toolathlon). These figures have not been independently verified.

Deployment Footprint

At BF16 precision, the model's weights alone would require roughly 4.8 TB of memory — more than an 8-GPU node can hold. Community-produced NVFP4 (W4A4) quantization compresses this to approximately 1.2 TB, fitting within a single node equipped with 8× NVIDIA B300 Blackwell Ultra GPUs (2.1 TB aggregate HBM3e). AWS has published a deployment guide showing the model running on Amazon SageMaker HyperPod using an ml.p6-b300.48xlarge instance with vLLM, covering NVFP4 quantization, tool calling, and MTP-based speculative decoding through an OpenAI-compatible endpoint.

No official hosted API pricing has been disclosed for Qwen3.8-Max or the open-weight release; organizations self-hosting the model avoid per-token fees but must provision purpose-built GPU infrastructure.

What This Means

Qwen3.8-2.4T-A95B extends the trend of frontier-scale MoE models being released with open weights, following similar moves with models like Kimi K3. For teams building agentic coding or long-horizon research systems, the appeal is control: data stays in-house, and the hybrid linear/full-attention design keeps memory bounded even at million-token context lengths. The catch is that only well-resourced teams — those with access to multi-GPU Blackwell-class nodes and orchestration platforms like SageMaker HyperPod — can realistically self-host a model of this size, even after 4-bit quantization. This positions Qwen3.8 less as a broadly accessible open-source alternative and more as a frontier-class option specifically for organizations with serious infrastructure budgets.