Alibaba Releases Qwen3.8-Max Open Model Weights
Alibaba has released the weights for Qwen3.8-2.4T-A95B, the post-trained model underlying its Qwen3.8-Max flagship, through Hugging Face. The mixture-of-experts model has 2.4 trillion total parameters but activates 95 billion per token, and supports up to 1 million tokens of context. Its custom license introduces commercial restrictions for some large model-service providers, according to a Yahoo Finance-published analysis.
Alibaba has released the weights and configuration files for Qwen3.8-2.4T-A95B, the post-trained model underlying its Qwen3.8-Max flagship, on Hugging Face. The release makes Alibaba's largest open-weight Qwen model available in the Transformers format, with stated compatibility for inference systems including vLLM, SGLang, and TokenSpeed.
The model has 2.4 trillion total parameters, with 95 billion activated parameters per token, according to its Hugging Face model card. It uses a fine-grained mixture-of-experts architecture and a hybrid attention design. NVIDIA describes the architecture as combining full and linear attention, with a context window of up to 1 million tokens and output length of up to 128,000 tokens.
Alibaba's model card describes Qwen3.8 as the most capable generation in its open-model family and lists configurable reasoning controls, including reasoning_effort and preserve_thinking. The latter retains reasoning context from earlier messages, according to the documentation. The company also distinguishes the downloadable A95B checkpoint from the managed Qwen3.8-Max service: the model card states that the official API version adds features including vision input, non-thinking mode, built-in tools, and a 1 million-token default context length.
Deployment remains a data-center problem
Open weights do not make a 2.4 trillion-parameter model lightweight to operate. NVIDIA writes that serving the model requires data-center-scale accelerated compute and distributed, multinode inference infrastructure. In an NVIDIA GB300 NVL72 configuration using FP8 precision, NVIDIA reported more than 4,000 tokens per second per GPU and more than 350 tokens per second per user without additional tuning.
For ML platform teams, the active-parameter count is relevant to per-token compute, but total parameter count still affects model storage, memory planning, loading, expert placement, and distributed serving complexity. Companies operating comparable large MoE checkpoints generally need to evaluate expert parallelism, interconnect bandwidth, quantization behavior, batch sizing, and tail latency rather than treating active parameters as a full proxy for deployment cost.
Artificial Analysis ranks Qwen3.8-2.4T-A95B second among 106 models on its Intelligence Index, with a score of 58. Its published evaluation also reports 46.3 output tokens per second and a 984,000-token context window. The organization characterizes the model as relatively expensive and slow compared with other open-weight models in its comparison set, reporting listed API pricing of $2 per million input tokens and $6 per million output tokens. Those measurements are third-party benchmark and pricing observations, not Alibaba performance claims.
License introduces a commercial threshold
The Hugging Face repository identifies the model license as qwen3.8-max, rather than an Apache 2.0 license. A Yahoo Finance-published article by Ethoswarm reports that the custom terms require a separate commercial license for Model-as-a-Service or AI Work Assistant businesses with aggregate revenue above $50 million across any consecutive 12-month period.
That reported threshold matters for organizations considering hosted inference or fine-tuning services built around the weights. Teams should review the full license text, rather than infer rights from the availability of the checkpoint. In the wider open-weight ecosystem, licensing terms can affect procurement, downstream redistribution, customer contracts, and whether a model is suitable for a managed-service business even when technical evaluation results are strong.
The release follows Alibaba's earlier public preview of Qwen3.8-Max. The Decoder reported on August 3 that weights were forthcoming and described demonstrations involving autonomous coding, research reproduction, and simulated e-commerce operations. Those demonstrations are company-provided examples, while the newly available weights allow developers to independently test the A95B checkpoint in their own agent, coding, and long-context workloads.
Key Points
- 1Alibaba released a 2.4 trillion-parameter open-weight checkpoint, expanding access to frontier-scale MoE experimentation while leaving deployment technically demanding.
- 2The model activates 95 billion parameters per token, but total weights still impose major memory, storage, networking, and distributed-serving requirements.
- 3A reported $50 million commercial-service licensing threshold makes legal review as important as benchmark testing for hosted model providers.
Scoring Rationale
The release exposes weights for a frontier-scale 2.4 trillion-parameter Qwen checkpoint, making it highly relevant to teams developing open-model and agentic AI systems. Its scale limits practical self-hosting to well-resourced operators, while the custom commercial license may constrain some service-provider use cases.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
