FC
Fern Cascade
14 hours ago
TurboFieldfare on M2 Air 8GB: 5.4 tok/s, then Metal hung after sleep
TurboFieldfare
Gemma 4 26B-A4B inference runtime for Apple Silicon Macs
finally got TurboFieldfare building on my M2 Air (8GB). the 15GB model pull took forever on cafe wifi, but once it streamed experts off SSD i was seeing ~5.4 tok/s — honestly usable for local draft work.
one weird bit: first run after sleep the Metal path hung until i relaunched the decode service. anyone else hitting that, or is it just me being on macOS 26 beta?
2 comments
Join the discussion
Log in to comment.
same hang on an M3 Pro after lid close. relaunching the decode service fixed it for me too. curious if they pin a Metal command buffer across sleep — llama.cpp never did this to me on the same machine.
did you build from main or a release tag?
speed checks out vs their 5.1–6.3 on 8GB Air. more interested in quality though — did you run even a tiny prompt set against mlx Gemma?
i tried three coding prompts last night and got solid completions, but one kept inventing a nonexistent Swift API. tok/s is nice; silent wrongness still costs more time.