My Friday night somehow turned into “let’s see if Codex can get Gopher64 running on macOS” and now Mario 64 is booting on my Mac like it was always meant to be there. Honestly, that’s a pretty good night.
@OpenAIDevs
English
Patrice Dery
54 posts
Patrice Dery
@PowerDryX
Building local-first AI tools on Mac. Creator of QwenVoice. Also making EXPELLED!! in Godot.
Katılım Aralık 2025
224 Takip Edilen6 Takipçiler
@HuggingModels I made a macOS app around this model family called QwenVoice for anyone who wants to try Qwen TTS/custom voices without having to deal with setup or config.
github.com/PowerBeef/Qwen…
English
Ever wanted to clone a voice from just a few seconds of audio? Meet Qwen3-TTS, a text-to-speech model that can mimic voices with incredible accuracy. It's blowing up with nearly 1M downloads because it makes custom voice creation accessible to everyone.
English
Patrice Dery retweetledi
Patrice Dery retweetledi
24h after the launch of Remodex: Codex Remote Control
Stats:
→ 1,354 Downloads
→ $1,220 Total Revenue
→ $553 MRR
→ 149 paying users
I have no words.
In 24h I made the app that made more than my past 10 projects combined together...
I'm so thankful to you guys
Thank you for the trust, and for the support
More to come.
Emanuele Di Pietro@emanueledpt
Introducing Remodex ✨ The first Codex Remote Control for iOS! Codex runs on your Mac. Your iPhone controls it, from wherever you are. Pair with one QR scan: → Create threads → Use subagents & skills → E2EE → Git actions and more Open source. Live now on the AppStore!
English
QwenVoice 1.2 is out.
This release is a big one for speed, stability, and overall polish.
It’s still the same fully offline Qwen3-TTS app for macOS on Apple Silicon, but it now feels a lot snappier. In my own testing on a base M2 Mac mini (8 GB), typical short voice clips went from 30s+ to under 5s.
1.2 also brings a refreshed UI, a cleaner generation workflow, live streaming preview for single generations, and a dedicated Voice Design workflow instead of hiding it inside Custom Voice. Saved voices and history got some love too.
Huge thanks to the open-source projects QwenVoice builds on — Qwen3-TTS, MLX, mlx-audio, and GRDB.swift — and especially to @Alibaba_Qwen, @Prince_Canuma, and @groue.
Grab it here and tell me how it feels on your Mac:
github.com/PowerBeef/Qwen…
English
@SaviourOfSoil It's live now! Grab it from the releases page and let me know if you run into any issues.
English
QwenVoice update: I'm officially done trying to hack emotion/tone control into Qwen3-TTS voice cloning 😅
I've explored every workaround I could find and none of them really hold up. The Base model that handles cloning just doesn't have an instruct interface — it's not a bug, it's the architecture.
Good news: the Qwen team has a VoiceEditing model in the technical report that combines cloning + delivery instructions. It's just… not released yet.
So I'm parking this feature and focusing on other parts of QwenVoice until a capable model drops. Sometimes the right move is to wait.
github.com/PowerBeef/Qwen…
English
That's exactly what 1.2 fixes — you can now save a voice you like from Voice Design and reuse it in the Voice Clone tab. Same voice every time, no more variation between generations.
Only tradeoff: Voice Clone doesn't support tone/emotion instructions, so you can't steer delivery. But the voice itself stays perfectly consistent.
English
@PowerDryX yes, same prompt and you have voice tone everything so different in just next generation
English
Patrice Dery retweetledi
i am not so smart about this but this should be good for edge AI too right
Prince Canuma@Prince_Canuma
Just implemented Google’s TurboQuant in MLX and the results are wild! Needle-in-a-haystack using Qwen3.5-35B-A3B across 8.5K, 32.7K, and 64.2K context lengths: → 6/6 exact match at every quant level → TurboQuant 2.5-bit: 4.9x smaller KV cache → TurboQuant 3.5-bit: 3.8x smaller KV cache The best part: Zero accuracy loss compared to full KV cache.
English
That’s basically the difference.
Ryan / Aiden / Serena / Vivian are already baked-in voice identities in the model, so they’re naturally much more consistent from one generation to the next.
A custom prompt-made voice is being recreated from a description each time, so even with the same prompt, the model can still land on a slightly different voice/tone on different runs.
Once a designed voice is saved and reused as a fixed reference, it gets much closer to that built-in voice behavior.
English
@PowerDryX where is the voice for ryan, aiden, serena, vivian coming from? if we can use custom prompt create a voice and save it, we would be able to use it just like ryan, aiden, serena, vivian...It will do the job!
Right, now each time we get different tone and voice via same prompt.
English
Yeah, this is feasible, and it’s going into the imminent 1.2 update.
I’m adding saved prompt-made voices via voice clone, so instead of rebuilding the same voice every time, you’ll be able to save it once and reuse it later.
Only caveat: tone/delivery instructions not carrying over in voice clone is a current model limitation, not a QwenVoice limitation. So this is mainly about preserving the voice itself, not all of the original expressive behavior.
English
@PowerDryX we dont need that, if somehow we can save custom made voice (the one we made using the prompt) and use that instead!
English
QwenVoice 1.2 is coming soon.
Big performance improvements this time. Voice clips now generate almost instantly, even on my 8 GB M2 Mac Mini. The live streaming preview feels buttery smooth and the whole experience is much more responsive.
This release also bumps mlx-audio to v0.4.1. Huge thanks to @Prince_Canuma for the outstanding work on the library.
The UI is now cleaner and more polished. Full release and changelog dropping very soon.
Stay tuned.
English
Patrice Dery retweetledi
🚀 mlx-vlm v0.4.1 is out!
What’s new:
🧠 New models — Mistral4, Molmo Point
🔧 Server improvements — model & adapter preloading, common sampling params, cleaner defaults
🛠️ Fixes — Qwen3-Omni integration, Qwen3.5 continuous batching
📋 Tool call index per OpenAI spec
⚡ Torch removed as a dependency
Welcome new contributors @auggie246 & @howeirdo! 🎉
Get started today:
uv pip install -U mlx-vlm
If you love it, leave us a star ⭐️
github.com/Blaizzy/mlx-vlm
English
It looks very cool but I’ll stick to Qwen3-TTS for QwenVoice for now. The whole idea behind the app is to be easy to use out of the box and well optimized for Apple Silicon — supporting multiple models would kind of defeat that purpose. Still working hard on the next release though! 🙏
English
@PowerDryX @PowerDryX if we can make this run in m series, then THIS IS IT!
English
Trying to make the app feel cleaner, more focused, and more like a real native macOS creative tool instead of “just an AI app UI.”
Still iterating, but I’m pretty happy with the direction so far.
Would love honest feedback on the layout, clarity, and overall feel.
English
@SaviourOfSoil You’re right and honestly, the UI direction became the main roadblock for the next release. I’m moving back toward a more native macOS look and keeping things cleaner and more focused. Thanks for the honest feedback.
English
@PowerDryX cool...just a little feedback...native macos looks were better than this gradient background
English
@SaviourOfSoil Not yet — I’m hoping to get it out this weekend 🙂 Really appreciate all the feedback. This update brings much better stability, a major performance boost, a fully redesigned UI, and some of the workflow improvements you asked for are in there too.
English
feature request for @OpenAIDevs:
> plan in ChatGPT w/ GPT-5.4 Pro
> push plan to Codex app
> Codex app breaks the plan down into tasks and builds
you can just build things.
i'd like to just build things even faster.
English
Patrice Dery retweetledi
Building mlx-audio with @lllucas and the community over the past couple years has been incredible.
Not for clout. Just an itch to scratch, guided by one motto: "helping people through technology."
Watching the community run with it has been the best part.
AVB@neural_avb
Been playing with mlx-audio for a side project on my mac. It is absolutely insane. Great work. You can do a lot with this. 💙 - text-to-speech, voice clone, customization ex: qwen-3-tts to kokoro to soprano - speech to text ex: whisper, parakeet (i will use mlx_whisper for whisper models though. that one supports word-level timestamps.) - speech to speech (noise removal, enhancement) ex: sam-audio, mossformer2 - supports the best open source audio models from the mlx-community The mlx-community on HF is so active as well, MAC is catching up atleast when it comes to running inference with the best open-source models.
English
Patrice Dery retweetledi
Starter Kit: Racing for @godotengine is now available! 🎉 Easy to understand code, fully open-source & 3D models included! It's a little rough around the edges, but feel free to do pull requests if you can improve on it!
github.com/KenneyNL/Start…
#gamedev #gameassets
English