Qwen38 Megakernel now refuses multiple agent slots on a group of NVIDIA graphics cards where the previous release could silently return wrong tokens. One slot remains usable. The developer says the new refusal could not be tested on affected hardware. Separate corrections restrict video inputs and remove an access key from a properties response.
Qwen38 Megakernel, a local engine and server for the Qwen3.8-27B language model, now prevents a configuration associated with incorrect output. The developer says graphics cards with 50–79 streaming multiprocessors could previously start multiple agent slots despite insufficient computation blocks for the batched path. The November 1 release refuses that combination during startup. Running a single agent slot remains permitted on those cards.[1]
The issue could yield incorrect tokens, the text pieces generated by a model, without an error message. Cards with 80 or more multiprocessors were outside the affected group. The developer expressly says an affected device was unavailable for testing the new startup refusal. The project also reports unchanged graphics-processing kernels and matching output in its own comparisons against the previous release.[1]
Video uploads are now limited to selected media containers, closing a path through media manifests to local-file reads or sender-chosen network requests. The notes say a server listening locally by default, or protected with an application programming interface key, was not exposed to that path from other machines. The properties endpoint also stops showing the key, and requests for zero output tokens are rejected.[1]
