Alibaba has published the weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open‑weight model to date, bringing near‑frontier capabilities to the open ecosystem. The model contains 2.4 trillion (2.4T) total parameters, with 95 billion (95B) parameters activated per token.
Architecture and intended use cases
Qwen3.8-2.4T-A95B uses a fine‑grained Mixture‑of‑Experts (MoE) design combined with a hybrid of full and linear attention layers. It supports contexts up to one million tokens and output lengths up to 128K, targeting demanding agentic workloads such as coding, large‑scale document analysis and long, multi‑step workflows.
The hybrid attention alternates between full‑attention layers—where every token attends to every other token—and linear‑attention layers, where a bounded recurrent state replaces a growing KV cache. This combination keeps compute and memory bounded as context sizes scale toward one million tokens.
Fine‑grained MoE distributes capacity across a larger number of smaller experts rather than a few large ones. A learned router activates only the experts needed per token, so serving costs follow the activated parameters and not the full 2.4T parameter count. This design aims to provide frontier‑scale capacity at a fraction of the cost of an equivalent dense model.
The model also includes built‑in reasoning controls (low/high/xhigh) that let developers trade compute for deeper reasoning quality on a per‑request basis—raising the setting for complex multi‑step reasoning or lowering it for high‑throughput document processing.
Multinode performance: GB300 NVL72
Running a 2.4T parameter open‑weight model requires data‑center scale accelerated compute and close co‑design of chips, systems and software. NVIDIA is collaborating with the open‑source ecosystem to enable multinode deployments via optimized kernels, inference runtimes and distributed serving recipes.
NVIDIA’s GB300 NVL72 integrates 72 NVIDIA Blackwell Ultra GPUs into a single rack‑scale platform. The large 72‑GPU NVLink domain provides efficient all‑to‑all communication at 130 TB/s, which alleviates bottlenecks that can occur when expert traffic must cross standard off‑the‑shelf networks.
On Day‑0, without additional model tuning, Qwen3.8-2.4T-A95B running in FP8 precision on a GB300 NVL72 achieves over 4,000 tokens per second per GPU and over 350 tokens per second per user. NVIDIA expects further optimizations, including NVFP4 precision, to yield additional performance improvements over time.
Serving and fine‑tuning options
NVIDIA supports multiple inference stacks to address different developer needs. SGLang, vLLM and NVIDIA Dynamo provide open‑source inference recipes for developers who want granular control over performance on NVIDIA‑accelerated platforms. The model can also be deployed via a model‑free NVIDIA NIM, a single inference container that serves any supported model.
Developers can post‑train the model for domain‑specific tasks using NVIDIA NeMo AutoModel, a PyTorch‑native fine‑tuning library with Day‑0 Hugging Face checkpoint support. NeMo AutoModel allows direct training on existing checkpoints without conversion and supports full SFT or memory‑efficient LoRA fine‑tuning.
Where to get the model
Qwen3.8-2.4T-A95B model weights are available for download from Hugging Face or ModelScope. Deployment can be performed with the model‑free NVIDIA NIM container available from NVIDIA NGC to serve fine‑tuned checkpoints and scale to production.
In summary, Qwen3.8-2.4T-A95B introduces a large, MoE‑based open model optimized for very long contexts and agentic workloads. Operating it requires multinode, data‑center resources, and NVIDIA reports significant Day‑0 throughput on GB300 NVL72 hardware.
