“MLX”

Thank you for the response and for pointing me toward MLX — I’ll definitely dive deeper into it. I understand the limitations around sharing roadmap details, and I’m looking forward to WWDC25 for any related announcements. Thanks again! — Aishwarya Dekhane

Thanks for the tip! I'll have to try MLX out later. Right now I need some results. I'm taking a course and a full week behind because all the data and models that I generated in the previous week is now suspect.

Hello, We don't offer something that specific so we're wondering if you've considered the Tensorflow Plugin for Metal. We also offer Metal-cpp if you prefer pure C++ to Obj-C++. Another framework you may want to consider for ML is MLX.

If your goal is inference, Apple Silicon with unified memory sidesteps these driver issues entirely. I've been loading 30B+ models via MLX on an M2 Pro — no PCIe bottleneck, no VRAM split, no driver compatibility fights. Might be worth comparing the cost of a Mac Studio vs the time spent debugging ROCm on the 2019 Pro.

Hello, We need some more details to help us understand your questions. Is mps introduced website a reference to Metal Performance Shaders? With respect to PyTorch, are you asking if MPS-accelerated PyTorch has been updated to use Metal 4? Separately, have you considered MLX for machine learning optimization?

Hi, Thank you for pointing me out to MLX, it's definitely a good option for developing new projects. However, PyArrow is a dependency for several packages. I went over the error, and there is a bug in the backend cpp code. I am not sure if you want me to report that here: Feedback Assistant.

Thanks for asking about this. I tried adding MLX Swift as a package to an app playground project in Swift Playground. I could not find a way to get it installed. None of the versions I tried could be added. Please file a report with Feedback Assistant: https://feedbackassistant.apple.com Please post the Feedback number here so I can check it.

Still broken as of early 2026 in my testing. For JAX workloads on Apple Silicon I've moved to MLX entirely — the API is different but the Metal backend actually works and gets regular updates. For anything that must stay in JAX, CPU fallback is unfortunately the only reliable path on macOS right now.

Tokenizer breakage across mlx versions is a recurring pain point — the tokenizer factory gets updated without guaranteed backward compat for custom-fused models. Check if tokenizer_config.json in your fused model specifies a tokenizer_class that 2.29.1 still recognizes. Manually setting the tokenizer type in LLMModelFactory registration usually gets around it.

I have M2 Max and faced similar issues ( not resolved fully yet ). but thing that helped me progress was different version of tensorflow and pythons ( generally 3.13.2) performed best , and tensorflow 2.16 give a shot , or only MLX then is option :-( pip install tensorflow-macos 2.16 Downside is when you try coremltools it won’t convert to .mlmodels

I'm also interested in this. Apple Silicon is an exciting hardware platform for ML and we need guidance from Apple how it envisions the future ecosystem. Is there going to be more openness and support of the major ML frameworks (incl. JAX), or is jax-metal an abandoned tech demo and we're all meant to use MLX more or less exclusively? The latter scenario would be incredibly unfortunate.

Solid work — 12-14 tok/s on A-series with 4-bit is respectable. 341-392 MB resident on 7.4 GB does leave thin margins though. Have you profiled whether MLX is placing any matmuls on ANE, or is this pure GPU? In my experience with Whisper-scale models the GPU path is more predictable but ANE helps with battery if the ops map cleanly.

I was surprised not to see a local model option after installing the Xcode 27 beta. I haven't watched any of the coding assistant videos yet, (very much not interested in cloud-based AI providers) but I am also curious about whether MLX has been found to be just not as good as the cloud-based providers, or whether there is some other reason Apple decided not to include a local-model option.

Thanks for these replies! I'll try them both today or tomorrow. I was able to get a fresh install to work on a different MacBook Air M2. Although simple tf/keras/gpu test scripts required several edits to even run and display versions and available devices! I've seen a comment on yt that the reason I can't find answers is because M3 users likely prefer the apple mlx libraries over tensorflow due to unified memory support. Thoughts?

To fully utilize local models in Xcode 27 beta 6 you need three things: A MLX model from 🤗Hugging Face (https://huggingface.co/models?library=mlx&sort=trending) An inference server, I'm using oLMX (https://omlx.ai) A tool using agent, I use OpenCode (https://opencode.ai) There is lots of info on how to configure OpenCode to use oLMX IMPORTANT Xcode 27 Beta 5 and later requires that the Agent executable be signed or Xcode crashes. Every time you update OpenCode you will need to resign the new executable Self Sign OpenCode # locate you opencode which opencode # codesign it codesign --force -s - /Users/gdunham/.opencode/bin/opencode # verify it codesign --verify --verbose /Users/gdunham/.opencode/bin/opencode # later update opencode opencode upgrade Multiple Models with OpenCode The way to use multiple models with OpenCode in Xcode 27, is to create a custom agent for each model and use environmental variables to specify the model.