Software Dev 👨💻 | Computer Eng. | YOLO 🚀 | AI ✨ | Tweets = personal. RTs ≠ endorsements ⚡ Mobile Inference : https://xitter.cf/t.co/jytHWegVUQ
United Kingdom
Joined March 2009
- Tweets791
- Following376
- Followers225
- Likes1.5K
Pinned Tweet
System update complete 🚀
2× RTX PRO 6000 Blackwell + 2× RTX 5090.
384GB total VRAM.
xitter.cf/cguldogan/status/20340…
FLUX 3 Image Klein'ı mflux'a port ettim ve 0.23.0 sürümüyle yayında. Apple Silicon'da MLX ile yerel olarak çalıştırabilirsiniz: metinden görsel üretimi, 10'a kadar referans görselle düzenleme, 4/8-bit ağırlıklar ve LoRA desteği.
Bağlantılar aşağıda 👇
Kurulum:
uv tool install --upgrade mflux
mflux: github.com/mflux-community/m…
4-bit ağırlıklar: huggingface.co/cguldogan/flu…
8-bit ağırlıklar: huggingface.co/cguldogan/flu…
I ported FLUX 3 Image Klein to mflux, and it's out in 0.23.0. Run it locally on Apple Silicon with MLX: text-to-image, editing with up to 10 reference images, 4/8-bit weights and LoRA support.
Links below 👇
Install:
uv tool install --upgrade mflux
mflux: github.com/mflux-community/m…
4-bit weights: huggingface.co/cguldogan/flu…
8-bit weights: huggingface.co/cguldogan/flu…
Got Fish Audio's Drama 3 (35B MoE text-to-speech) running natively on Apple Silicon with MLX 🎙️ Voice cloning, up to 16 speakers, inline [laughs]/[sighs]. ~1.4× faster than real time on an M4 Pro. Built with Fish Audio.
Two tiers:
4-bit (24 GB, needs a 48 GB+ Mac) and
8-bit (40 GB, 64 GB Mac).
Validated against the official PyTorch implementation; it handles English and Turkish cleanly.
Code and a local test UI (voices, speaker tags, takes and edits)
Kod ve yerel test arayüzü (sesler, konuşmacı etiketleri, kayıtlar, düzenleme):
github.com/cguldogan/mlx-aud…
Three variants:
• 8bit 5.7 GB, ~40 tok/s, near-identical to bf16
• 4bit 3.5 GB, ~64 tok/s
• bf16 9.9 GB, the reference
The encoder stays fp32: its activations reach ~250k, which overflows fp16.
Built on @Prince_Canuma's mlx-vlm, which made this possible. Upstream PR is open, so hopefully it's a plain pip install soon:
github.com/Blaizzy/mlx-vlm/p…
My first model port and my first quantization 🎉
Tencent's Youtu-Parsing-Omni (5.3B; documents, charts, audio and video → structured JSON) now runs on Apple Silicon with MLX.
8bit: 5.7 GB, ~40 tok/s on an M4 Pro
It’s crazy !!
I gave Claude Code a phone number. 📞
It calls me when a task is done, it's stuck, or it needs a decision. And I can call it back and tell it what to do.
All self-hosted, open source. 🧵
Install it with one prompt pasted into Claude Code:
github.com/cguldogan/on-phon…
