Midjourney releases V8 Alpha
The company said it rewrote its codebase from Google TPU infrastructure to a GPU-native PyTorch stack, enabling native 2K output at roughly five times the generation speed of V7.
- Models & capabilities
- Minor
Midjourney released V8 Alpha in early access, describing it as an entirely new model built on infrastructure the company had rewritten from scratch — moving from its previous TPU-based setup to a GPU-native codebase built on PyTorch. The company said years of accumulated workarounds on the old stack had made new features increasingly difficult to build, and that the rewrite removed those constraints.
The immediate, user-facing gains were speed and resolution: Midjourney said image generation ran roughly five times faster than under V7, and introduced an --hd mode that renders natively at 2K resolution rather than upscaling a lower-resolution output. A --q 4 setting offered higher coherence at four times the cost and generation time. The release also added new personalisation tools, including style references and moodboards, and new prompt parameters (--chaos, --weird, --exp, --raw), while remaining backward-compatible with users’ existing V7 personalisation profiles. Midjourney said its cheaper “relax” generation tier was not yet supported on the new architecture and that a lower-cost option was in development.
Midjourney has continued to iterate quickly on image models through alpha and point releases rather than infrequent major versions, a cadence it credited in this case to the new GPU-native foundation making features like HD mode and faster personalisation newly practical to build.