Qwen3.8-Flash-Next on a 64 GB M2 Ultra: A 66-Minute Real Work Run (b1tank.github.io)

Qwen3.8-Flash-Next has demonstrated its practical application for real-world tasks on the 64 GB Apple M2 Ultra, successfully completing a sustained 66-minute coding session with 106 tool calls and an average generation rate of 35.5 tokens per second. This session involved debugging and packaging a local-first OpenTelemetry tool called OTelux for macOS and showcased the model's capability to handle extensive coding tasks without monopolizing system resources. Notably, the model effectively compacted conversation history and resumed work even after encountering issues, suggesting its robustness in real-world scenarios. This performance is significant for the AI/ML community as it indicates that large language models can operate efficiently even on consumer-grade hardware, increasing accessibility for developers. The use of mixed precision and selective disk reads during runtime allowed it to maintain a manageable memory footprint, which is crucial for running complex applications. The session's success, including a seamless execution of long-context recovery and essential debugging operations, underscores the practical implications of advanced AI tools for software development, potentially transforming workflows in coding and debugging environments.