GPT-OSS 120B tops 60 tok/sec on M5 Max

The arrival of "workstation-class" performance on a laptop marks the end of cloud dependency for privacy-sensitive professional workflows.

• –MoE architecture only activates 5.1B parameters per token, allowing the 120B model to achieve throughput typical of much smaller dense models
• –M5 Max's 614 GB/s memory bandwidth is the critical enabler, effectively doubling the performance of prior generations for large-scale local inference
• –MXFP4 quantization preserves high precision while fitting the model within 70GB, leaving ample room for 128k context windows on 128GB machines
• –Apache 2.0 licensing combined with local hardware provides a viable, HIPAA-compliant alternative to proprietary APIs for clinical and legal document processing