M3
A new-generation large language model released by MiniMax, serving as the replacement for M2.5 and M2.7.
Timeline (last 90 days)
For new projects not yet using M2.5/M2.7, adopting M3 directly is the more reasonable choice, avoiding another migration in the short term.
The public materials do not provide specific pricing data for M3 or exact API change details.
MiniMax officially recommends that users of M2.5 and M2.7 migrate to the M3 model.
The base M3 has a memory bandwidth of about 100 GB/s.
The developer's desktop test devices include a MacBook with an M3 chip as the primary Apple Silicon test machine, and an older Intel MacBook running Omarchy (an Arch-based Linux configuration).
Apple's ANE neural engine was recently reverse-engineered, and someone has reported a 50% prefill performance improvement on the M3.
Consumer-grade M2/M3 MacBooks can run quantized models with parameter counts up to 70B locally at relatively low latency.
The M3's GPU introduces Dynamic Caching, enabling hardware-level on-demand resource allocation.
The M3 is manufactured using TSMC's first-generation 3nm process (N3B).
GPU drivers written for the M1 and M2 cannot be directly reused for the M3 and require re-adaptation.
4 more timeline events
All Facts (14)
For new projects not yet using M2.5/M2.7, adopting M3 directly is the more reasonable choice, avoiding another migration in the short term.
50%UnverifiedThe public materials do not provide specific pricing data for M3 or exact API change details.
50%UnverifiedMiniMax officially recommends that users of M2.5 and M2.7 migrate to the M3 model.
50%UnverifiedThe base M3 has a memory bandwidth of about 100 GB/s.
50%UnverifiedThe developer's desktop test devices include a MacBook with an M3 chip as the primary Apple Silicon test machine, and an older Intel MacBook running Omarchy (an Arch-based Linux configuration).
50%UnverifiedApple's ANE neural engine was recently reverse-engineered, and someone has reported a 50% prefill performance improvement on the M3.
50%UnverifiedConsumer-grade M2/M3 MacBooks can run quantized models with parameter counts up to 70B locally at relatively low latency.
50%UnverifiedThe M3 is manufactured using TSMC's first-generation 3nm process (N3B).
50%UnverifiedThe M3's GPU introduces Dynamic Caching, enabling hardware-level on-demand resource allocation.
50%UnverifiedAsahi Linux on the M3 currently has limitations including incomplete GPU acceleration, limited peripheral support, power management pending improvement, and external display compatibility issues.
50%UnverifiedThe M3 is manufactured using TSMC's first-generation 3nm process (N3B) and contains approximately 25 billion transistors.
50%UnverifiedGPU drivers written for the M1 and M2 cannot be directly reused for the M3 and require re-adaptation.
50%UnverifiedThe M3 series introduced the Dynamic Caching feature, enabling hardware-level on-demand allocation of GPU resources.
50%UnverifiedThe Asahi Linux project announced that it is officially extending support to Mac devices equipped with the M3 chip.
50%