@JoshuaSWarren

Omarchy M team. Making Apple Silicon fast under Linux: MLX, Mesa GPU, Neural Engine. Shipped an iPhone app with no Mac. Building Remnic.

 Dallas, Texas
joshuawarren.com/sponsors
Joined May 2011
  • Tweets16K
  • Following6.2K
  • Followers7.8K
Apple locks fast local AI to macOS with Metal. I picked the lock. metal2vk translates Metal GPU code to Vulkan, unchanged. 598 of 610 real programs from @trymirai's uzu and @ashxhart's TensorFold so far. Engines built for Macs now run on Linux. github.com/joshuaswarren/met…

GitHub - joshuaswarren/metal2vk: Compile Metal Shading Language compute kernels to Vulkan SPIR-V...

Compile Metal Shading Language compute kernels to Vulkan SPIR-V (clang C++ for OpenCL + a metal_stdlib shim + clspv) - joshuaswarren/metal2vk

github.com
31
57
10
641
44,761
Here’s a thin Omarchy laptop doing AI photo masks with Mac Studio power. LightCraft on Linux now runs SAM3 on a remote Mac over SSH. You click on the laptop and the big machine does the inference. Merged in LightCraft #359. Same idea now PR’d to Rawmakase in #480.
1
19
631
A 35B model on a MacBook's GPU under Linux just got 1.67x faster per token. One kernel change in MLX: Qwen3.6-35B-A3B decode on an M2 Max went from 83 ms to 50 ms per token. Output is bit-identical. Built for @omarchy. Open source: github.com/joshuaswarren/oma…

gather_qmm decode: bind scales, biases and index arrays in place (gemma-4-26B decode 1.26x on M2...

Decode gather_qmm (routed-MoE experts at one row) copied every expert's scales and biases into a staging buffer on each call, to read a handful of experts. At 128 experts that is 15.9 MB of...

github.com
10
5
76
2,004
Huge thanks to @niels_gron, who sent lab hardware off my wishlist, starting with an AMD RX 9070, a Samsung 990 PRO, and Thunderbolt 5 and DisplayPort cables. It all goes into making local AI run on every GPU, on Linux Macs and beyond. Open source runs on people like Niels.
5
47
1,268
If anyone is wondering why this GPU - my vision is to unlock as much hardware as possible for open-source AI. Like many of you I don’t have the cash for newest NVIDIA cards. I can do the most good by unlocking support for less expensive hardware you may already have or can get.
1
7
199
And yes, once I have one of the eGPU docks off the list, I am going to get an AMD GPU running via an eGPU dock connected to Omarchy Linux running on Apple Silicon to do combined AI inference under mlx. A little crazy? Probably. Fun? Definitely!
4
138
AI on my 13-inch M1 under Linux was stuck at 10 tokens/sec. The fix was one setting: the governor. schedutil: 10.65 tok/s performance: 15.27 tok/s 1.43x faster decode. We're so early in this work, there's easy gains to be had still. Open source: github.com/joshuaswarren/oma…
3
1
1
68
1,961
To be precise: the 10 minutes was the agent building and notarizing on Linux. The App Store approval came after that (we all know that step takes them more than 10 minutes!), on the same Linux-built app. Mathias confirmed both above.
1
5
481
Joshua Warren retweeted
Mathias Beugnon@oomathiasOct 11
Replying to @ddalcu @JoshuaSWarren @omarchy
The build passed Apple Review. (not in 10 min)
1
1
1
1
224
Someone handed their agent my omarchy-apple-dev repo. 10 minutes later it had built and notarized a macOS app on a Linux machine. It passed Apple review. No Mac needed to ship to the Mac App Store. Built for @omarchy, open to everyone.
Mathias Beugnon@oomathiasOct 10
Replying to @JoshuaSWarren
Sol took 10 min to build and notarized a MacOS app on a Linux host; just by linking your repo. It passed Apple review. Amazing times.
22
24
3
466
13,209
Joshua Warren retweeted
Ash Hart
@ashxhartOct 10
What is your biggest pain point with local AI right now?
195
5
3
87
17,733
More on metal2vk, what it unlocks, and how to support the work: joshuawarren.com/projects/me…

metal2vk: Metal kernels on Vulkan | Joshua Warren

metal2vk compiles Apple Metal compute kernels to Vulkan, so they run on Apple Silicon under Linux with open-source tools.

joshuawarren.com
1
23
2,047
Omarchy on Apple Silicon is now merged into Omarchy core! One codebase no matter what architecture you run on. This will simplify getting my MLX on Linux work shipped. And we aren’t stopping here - M3 thru M6 support is all in the works. And crazy local AI performance boosts!
Ryan R. Hughes
@ryanrhughesOct 10
I've never been more excited to hit merge on a PR than I was for this one. 🎉 It took 15 people on five continents to get here. @jondkinney kept a branch alive for almost a year. @tiredkebab's fork pushed the idea further. Then came a lot of long conversations with @maralcbr and @bi9kahuna to get it over the line. Major props to the rest of the Omarchy M team who helped in various ways as well: @martiano @riverleaf88 @buildwithdjdev @RyanTheTide @Acelogic_ @LiamRay1O @ewninjaofficial @jaidipsubedi @JoshuaSWarren @shawnyeager @novuon_ai Thank you all. This one's special, and lays important groundwork.
There's a new version of this post
2
10
463
Some of you on non-Apple hardware wanted to run apple-dev without downloading Xcode. Made Xcode optional. omarchy-apple-dev builds Swift, UIKit & Flutter apps on Linux. They run on a real iPhone. Even with no Xcode download. No Apple SDK. Open source: github.com/joshuaswarren/oma…

GitHub - joshuaswarren/omarchy-apple-dev: Build and deploy iOS SwiftUI apps from Omarchy Linux on...

Build and deploy iOS SwiftUI apps from Omarchy Linux on Apple Silicon, no Xcode required - joshuaswarren/omarchy-apple-dev

github.com
22
40
5
514
20,463
Everyone says AI video needs a big NVIDIA GPU. This clip, sound and all, came from a 33B video model on a MacBook's GPU under Omarchy Linux. I ported @ashxhart's TensorFold to run it. Every step, text in to audio out, on an M2 Max. 59 minutes. Guide: github.com/joshuaswarren/oma…
6
1
38
2,246
more replies
It's not just video. On Omarchy Linux on Apple Silicon I also made: - llama.cpp up to 17x faster on the Mac GPU - MLX matmuls faster than macOS on the same M1 - iPhone apps built and debugged from Linux - A 753B model running across 5 machines
1
3
255
Next: a $7,500 DGX Spark, so one model can run across a Mac and a Spark. If this work saves you time or money, you can help fund it: GitHub Sponsors: github.com/sponsors/joshuasw… Buy Me a Coffee: buymeacoffee.com/joshuaswarr… Lab wishlist: amazon.com/hz/wishlist/ls/68…

Sponsor @joshuaswarren on GitHub Sponsors

I build open-source AI systems and developer tools: MLX on Apple GPUs under Linux, Apple Neural Engine support, and Remnic for agent memory. Omarchy M team member. Sponsor hardware testing, upstrea...

github.com
5
303
My favorite plugin I've shipped so far: shiplog, a live count of what you push and merge each day. 120, 146, 119, 198, 185, 125. That's PRs per day this week, most of them speeding up local AI and Apple dev on Omarchy Linux.
DHH
@dhhOct 9
Well over FIVE THOUSAND plugins now for Omarchy. And that's just the ones that have been published! Thousands more are truly bespoke and individualized creations just for the maker themselves. This is a whole new age of software for sure.
3
27
947
Load more