The promised release is now available, but downloading the Max-class weights does not reproduce every feature of the hosted Qwen3.8-Max service.
Read the exact model card
The official Qwen3.8-2.4T-A95B card describes released weights with 2.4 trillion total parameters and 95 billion active. They are text-only and require thinking. Native context is 262,144 tokens, with extension support; hosted Max offers additional features. The 27B card instead describes a dense vision-language model with optional non-thinking operation and Apache 2.0 licensing.
The large model uses a specific Qwen3.8-Max license. It includes display obligations at stated scale thresholds and a separate-license requirement for specified large model-service or AI-work-assistant businesses, with an internal-use exception. Do not substitute the 27B license for this one.
Active parameters are not the storage budget
The illustration shows nominal weight arithmetic, before runtime overhead. At two bytes per parameter, 2.4 trillion parameters represent 4.8 TB; at an ideal four bits, 1.2 TB. Using only the 95 billion active count would give 190 GB at two bytes, which does not account for the full expert set.
These decimal calculations are not deployment requirements or the sizes of published archives. Quantization metadata, tensors stored at other precisions, runtime buffers, context state and sharding overhead change the actual footprint. Offloading changes where weights live and how they move; it does not make the remaining experts disappear.
Keep API parity and task quality separate
Before replacing an endpoint, try each input type and control your application relies on: text, image, video, thinking mode and tool integration. Test the exact checkpoint and serving engine. A shared family name is insufficient to establish that those features match.
Then evaluate completed tasks under the same limits. The publisher’s benchmark tables describe particular models, harnesses and settings; they do not certify a self-hosted quantization with different tools. Record quality, elapsed time and total serving cost together.
For a local trial, 27B and the 2.4T mixture of experts are materially different capacity decisions. Begin with the required modalities, license and memory budget, then choose the artifact to evaluate. This is more reliable than treating a hosted product name as a complete specification for downloadable weights.
September 8: replace pending-weights claim with actual model cards; distinguish text-only thinking Max weights from hosted features and multimodal 27B, verify custom license and nominal memory arithmetic.