Alibaba's Qwen 3.8-Max Open Weights Arrive With a Catch: 2.4T MoE Is Text-Only at This Size

Alibaba kept its promise from the August 3 API launch: the open-weight checkpoint for Qwen 3.8-Max arrived on Hugging Face and ModelScope between August 12 and 13, 2026 under the model designation Qwen3.8-2.4T-A95B. The release makes the full 2.4-trillion-parameter sparse Mixture-of-Experts architecture publicly downloadable — but with important caveats that separate the open weights from the API product that launched last month.

The checkpoint is text-only. The multimodal capabilities — image, video, and document input — that make the Qwen 3.8-Max API competitive with Claude Opus 5 on vision tasks are not included in the public weights. Similarly, the 1-million-token context window in the API version is not reproduced in the open-weight checkpoint; context handling at that scale requires infrastructure that is part of Alibaba's serving stack, not the model weights alone. The architecture itself — 2.4T parameters, 95B active per token — is intact and verifiable.

The license is a custom agreement, not Apache 2.0. Acceptable-use terms impose restrictions on competing commercial deployments, meaning this is not a fully permissive open release in the way the Kimi K3 weights or Qwen 3.8-27B are under MIT/Apache licenses. Self-hosting the 2.4T checkpoint also requires substantial infrastructure — roughly 2TB of VRAM to serve the full model in standard precision — putting it beyond the reach of consumer or small-team deployment. Together AI and Modal are expected to host the checkpoint for API-based access to the raw open weights.

The nuance matters for the open-source AI narrative that has been building around each major Chinese lab release: the checkpoint is real, verifiable, and downloadable, but the gap between what you get from the open weights versus the Alibaba API is wide enough to matter for most practical use cases.