Alibaba Unveils Qwen3.8-A95B, a 2.4 Trillion Parameter Model
Designed for high-performance AI workloads on NVIDIA GB300 NVL72 platform.
1Massive-Scale Model
Alibaba has released the open weights for Qwen3.8-2.4T-A95B, a 2.4 trillion parameter open-source AI model. It uses a mixture of experts (MoE) architecture with 95 billion parameters activated per token. With a 1 million token context window and 128K output length, it's designed for complex reasoning and agentic workloads. It delivers high performance on the NVIDIA GB300 NVL72 platform.
2Technical Innovations
The model employs a hybrid combination of full attention and linear attention layers. In full attention layers, each token interacts with others, while linear attention layers use a bounded recurrent state. This approach keeps computation and memory usage under control even with growing context sizes. The fine-grained MoE architecture enhances specialization and reduces costs.
3NVIDIA Collaboration
NVIDIA is working with the open-source ecosystem to optimize multi-node deployments. Custom kernels, inference runtimes, and distributed serving recipes are being developed. On the GB300 NVL72 platform with FP8 precision, it processes over 4K tokens per second per GPU. Future optimizations like NVFP4 are expected to further boost performance.
4Designed for Agentic Workloads
Qwen3.8-2.4T-A95B is optimized for challenging agentic workloads like coding, large-scale document analysis, and multi-step workflows. Unlike traditional chat models, it can accumulate system instructions, tool outputs, and reasoning traces. Innovative architectural solutions maintain performance despite growing context sizes.
5Deployment Options
NVIDIA offers various inference stacks to meet developer needs. Open-source solutions like SGLang, vLLM, and NVIDIA Dynamo provide customizable performance. The model can also be deployed as a single container via NVIDIA NIM. Weights downloaded from Hugging Face and ModelScope are ready for immediate use.
6Infrastructure Requirements
Deploying a 2.4 trillion parameter model requires data-center-scale accelerated computing. The NVIDIA GB300 NVL72 integrates 72 Blackwell Ultra GPUs into a single platform. NVIDIA NVLink at 130 TB/s enables efficient communication for expert traffic, eliminating traditional network bottlenecks.
7Developer Tools
NVIDIA NeMo AutoModel provides a PyTorch-native fine-tuning library. Direct training is possible with Hugging Face checkpoints. Options include full SFT or memory-efficient LoRA fine-tuning. Developers can easily customize the model for domain-specific use cases.
8Industry Impact
Qwen3.8-2.4T-A95B brings near-frontier capabilities to the open-source AI ecosystem. It enables high-throughput production deployment of large-parameter models. AI factories benefit from low latency and high performance. This development could set new standards in the AI industry.