Got a pleasant surprize from DHL today, delivery date range was 22-25. After it was shipped from Apple delivery was set for Thursday, but this morning got an update that I should get it earlier and here it is =)
Specs:
M5U 36/80
256GB RAM
2TB SSD
YAY!
Those are out of the box, some1 with a preview access got the machine early and runs benches on oMLX which collects them and publishes.
Qwen27b 4b performance looks to be on par with rtx4090 (unsurprising considering specs).
>Qwen3.8-27B-MLX-4bit 4bit 8k 1,800 50.6
GLM-5.3 flash 4b numbers look quite bad tbh. I'm expecting twice that for ds4 flash (not 4.1) and qwen3.8-flash-next fp8.
>GLM-5.3-Flash-oQ4e 4bit 16k 955.4 33.0
edit:
>Qwen3.8-Flash-Next-oQ4e-mtp 4bit 64k 2,574 79.9
just appeared, looks not bad
edit2:
results are gone from the page for me now, there's a screen of them in the comments.
Privacy have been concern of many of us to have their own hardware to run llms, and here's another reason why: two mathematicians spent a year cracking one of the hardest problems in math and fed every draft of their works into Codex. A few days before they could publish, OpenAI suddenly showed up with the same solutions. When asked if their model (Sol and Astra) was trained on the pair's private chats, OpenAI did not answer the question.
Full statement from them https://cims.nyu.edu/~tristanb/statement.pdf
Feels like big labs believe everything you did with the help of their models is theirs.
This news seem to have flown under the radar. Apple announced CoreAI on WWDC which is basically a future replacement for CoreML and an alternative to MLX/llama.cpp/torch for on-device optimized inference, especially on phones and tablets.
The model weights need to be converted similarly to CoreML via python script, atm the list of supported models is mostly from mid 2025 year though https://github.com/apple/coreai-models/tree/main/models . For anyone wondering how is that anything new - CoreML out of the box didn't even support models beyond a few billion params and had very limited supported operations pool. This implies big update to ANE ops too.
There's nothing on performance yet, it is very likely that it's inferior to pure MLX on GPU atm. The only other interesting thing is that they boast 20B model to be deployed on device for foundation models https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models, which looks to be lazily loaded MoE, so perhaps CoreAI will allow to deploy larger models with apps as well.