Ömer Çelik

Engineer at Amazon

[Ömer Çelik @omercelik]

@TheAhmadOsman After Apple raised the minimal configuration to 512GB, second hand 256GB variant is now more expensive than the original selling price! 🙂

[Ömer Çelik @omercelik]

@sniperant Hangi donanim ustunde calisiyor?

[Ömer Çelik @omercelik]

Github: https://t.co/zp9xrHHmB5

[Ömer Çelik @omercelik]

🚀 Bir süredir üzerinde çalıştığım MLXtra adlı projem paylaşmaya hazır.

Mac için tamamen offline, tüm modeller bir arada: 🖼️ resim üretimi, 🎵 müzik üretimi, 💬 chat ve görsel anlama tek uygulamada.

Teknik kurulumlarla uğraşmadan, güçlü AI modellerini kolay bir arayüzle kullanabilir; sohbet, resim ve müzik üretimlerinizi verileriniz dışarı çıkmadan doğrudan kendi bilgisayarınızda yapabilirsiniz.

Deneyip yorumlarınızı ve önerilerinizi paylaşırsanız çok sevinirim 👇

[Ömer Çelik @omercelik]

@janekm @willowion It is open source, so you can check/adjust yourself 😊

[Ömer Çelik @omercelik]

@janekm @willowion You can use MLXtra to let Qwen or Gemma write it for you https://t.co/eSKxHSDehZ

[Ömer Çelik @omercelik]

Announcing MLXtra 📷

A native all-in-one local AI studio for Mac: image generation, music creation, chat + vision, and speech.

Runs locally on Apple Silicon, keeps your data on your device, and recommends models based on your Mac's hardware.

Built on Apple's MLX. Open source under Apache 2.0.

Details below 📷

[Ömer Çelik @omercelik]

@TeksEdge You can test locally using MLXtra https://t.co/eSKxHSDehZ

[Ömer Çelik @omercelik]

Announcing MLXtra 📷

A native all-in-one local AI studio for Mac: image generation, music creation, chat + vision, and speech.

Runs locally on Apple Silicon, keeps your data on your device, and recommends models based on your Mac's hardware.

Built on Apple's MLX. Open source under Apache 2.0.

Details below 📷

[Ömer Çelik @omercelik]

@mstrYoda_ Artik 256GB olan model satista olmadigi icin kar etmis olabilirler

[Ömer Çelik @omercelik]

I'd really appreciate a GitHub star for MLXtra ⭐

https://t.co/Xxsxl1xviJ

[Ömer Çelik @omercelik]

Announcing MLXtra 📷

A native all-in-one local AI studio for Mac: image generation, music creation, chat + vision, and speech.

Runs locally on Apple Silicon, keeps your data on your device, and recommends models based on your Mac's hardware.

Built on Apple's MLX. Open source under Apache 2.0.

Details below 📷

[Ömer Çelik @omercelik]

MLXtra is fully open source under Apache 2.0 and built for the MLX ecosystem on Apple Silicon.

Try it here → https://t.co/jVo0MtSxjk

Mac users: which would you try first: image generation, music creation, or vision chat?
Curious to hear your feedback.

[Ömer Çelik @omercelik]

Announcing MLXtra 📷

A native all-in-one local AI studio for Mac: image generation, music creation, chat + vision, and speech.

Runs locally on Apple Silicon, keeps your data on your device, and recommends models based on your Mac's hardware.

Built on Apple's MLX. Open source under Apache 2.0.

Details below 📷

[Ömer Çelik @omercelik]

Here's what makes MLXtra different:

One native Mac app
No model server setup, no terminal workflow, simple first run.

Local creative AI
Generate images with Ideogram 4, FLUX.2 Klein, or Z-Image Turbo. Create music with ACE-Step or Magenta RealTime 2.

Chat, vision, and speech
Use models like Gemma 4 and Qwen 3.6 for multimodal chat with built-in draft-model acceleration, plus Kokoro / KugelAudio for local speech.

Hardware-aware model picks
MLXtra recommends models based on your Mac’s hardware, so you don’t need to compare memory estimates manually.

Everything is designed to make local AI easier to use while keeping your data on your Mac.

[Ömer Çelik @omercelik]

@DrJesseMorse What is your MOTS-c and SS-31 protocol?

[Ömer Çelik @omercelik]

@grepmoney Thanks for sharing it! It gave the inspiration and I used the examples in replies to generate some similar training data. Will share it soon in @huggingface

[Ömer Çelik @omercelik]

I fine-tuned a Qwen3.5 9B model and cut thinking tokens by 57.40% and total tokens by 25.60%, with no visible quality drop in the actual response.

This suggests there is huge potential!

Next step: complete training, run proper evals, and share the quality/performance numbers.

[jeffypoo @grepmoney]

I just saw Codex leak a thinking trace that might explain why it is more token efficient. Small sample:

"Or if no org scope, keep legacy-only? But then write path not semantic. Could create report? No. Need ask user."

Codex thinks in grug brain to save tokens.

[Ömer Çelik @omercelik]

@onatm Benim kizim da soyle bir site yapti kendisine: https://t.co/a1LnVTNmo0

[Ömer Çelik @omercelik]

Training is done locally using mlx-tune. Thanks @_ARahim_

[Ömer Çelik @omercelik]

To compare the outputs, here is an example side by side!

[Ömer Çelik @omercelik]

Very efficient in thinking!

[Ömer Çelik @omercelik]

Another one - coding: