Ben Buschmann retweeted
Following my earlier post about the M5 Ultra 512GB Mac Studio, more than two dozen people asked me to set up crowdfunding for it.
I always try to pay my own way, so asking for help at this scale doesn't come naturally. But I couldn't make the numbers work this time, and my loan got rejected after placing the order.
I'd love to get one for TensorFold and to see how far we can push it.
If you'd like to help towards it, I've put the goal on Ko-fi below. Please only contribute if you're comfortable doing so.
ko-fi.com/ashxhart/goal?g=2
Thank you to everyone who's encouraged me to do this, and to everyone already supporting TensorFold. I'm incredibly grateful and will continue to push TF. ❤️
@TheAhmadOsman @0xSero @ivanfioravanti
Ash makes TensorFold for free but got turned down for a personal loan so can't buy a Mac M5 512GB to do more work on.
What do you guys think? You'd help to promote a crowd fund, wouldn't you?
We added support for Imagination Studio by Keys + Wasserman first (@SamJWasserman @u1tra_instinct). GapRunner now schedules GPU work it didn't start: an app asks for a reservation, waits its turn, and GapRunner makes room. The LLM sleeps; training parks at its last saved step.
Any app that starts its own GPU server can ask the same way: a small client wraps its start command, and GapRunner needs one job entry. GapRunner guide:
github.com/benbuschmann/gapr…
Imagination Studio: github.com/drowzeys/Imaginat…
I open-sourced GapRunner: a job queue for the GPU that runs your local model.
Training runs, evals and benchmarks wait their turn and run beside the model if they fit, in its idle gaps if not. A request comes in, they step aside, the model wakes.
github.com/benbuschmann/gapr…
Coding agents can run it too. There's an agent guide, a drop-in skill and llms.txt, so an agent can read the state, queue a job and explain why one is waiting.
One Python package, standard library only, Apache-2.0. Linux and one NVIDIA GPU. Early (0.x), in production on my Station since October. DGX Spark support is in testing.
It borrows from Berth and llama-swap, which are worth a look too.
github.com/benbuschmann/gapr…
1/
Benchmarked Apple's Private Cloud Compute model (AFM 3 Cloud) on macOS 27.
fm serve gives you an OpenAI-style endpoint on localhost, so this was curl and a Python timer.
Prompt processing: ~2,300 tok/s
Generation: 118–143 tok/s
A 26K-token essay summarized in 15s
3/
Catches:
• No quota numbers, just approaching_limit and limit_reached flags
• 32K context
• pcc only works if the server is launched from Terminal .app
• Tool calls come back as raw text, so Claude Code or Codex can't use it yet
Free, though. Good for summaries and rewrite
Early Darkbloom benchmarks (qwen3.6-35b-a3b-vl-mtp-mxfp8, decode tok/s):
🚀 Base M5 Max Studio: 122
💡 M6 Mini: 70
The M6 Mini beats my M1 Max (65) with less than half the memory bandwidth. The M5 Max is on another level.