Ignore my previous comment as I was testing the configuration on llama.cpp, which isn't as optimized as MLX based frameworks. However after trying this on mlx-vlm, I'm actually curious how you got this running long enough for your sanity testing. I tried to test it myself, and mlx-vlm is currently suffering from intense RAM spikes (4x the usage) which then very quickly causes OOMs. I opened an issue and I am working with collaborators who confirmed this is an issue.
Anthony A
anthonyya
AI & ML interests
None yet
Recent Activity
commentedon an article 5 days ago
Block Diffusion on Apple Silicon with 3.7× Speedup for Qwopus 3.6 27B liked a model 5 days ago
lemuralabs/Qwen3.6-27B-V2-abliterated-uncensored-6-bit-mlx updated a model 5 days ago
anthonyya/Qwen3.6-27B-DFlash-8bitOrganizations
None yet
commented on Block Diffusion on Apple Silicon with 3.7× Speedup for Qwopus 3.6 27B 5 days ago
liked a
model 5 days ago
updated a
model 5 days ago
published a
model 5 days ago
updated a
model 5 days ago
published a
model 5 days ago
updated a
model 5 days ago
published a
model 5 days ago
Are quants required to match?
#1 opened 6 days ago
by anthonyya
commented on Block Diffusion on Apple Silicon with 3.7× Speedup for Qwopus 3.6 27B 6 days ago
Can you run this on a real benchmark? I tested this on my M4 Max with a real coding benchmark and I'm not seeing anything close to your 55 tok/s, I'm getting around 20.
liked 2
models 13 days ago
liked a
Space about 1 year ago
liked a
model about 2 years ago
