Hey, I’m Gökdeniz Gülmez — an ML researcher and engineer. When I’m not building production systems, I’m deep in the MLX ecosystem, contributing to cutting-edge tooling for Apple Silicon.
I'm a contributor to mlx-vlm, mlx-examples, mlx-lm, and mlx, officially acknowledged by the MLX team — check here for mlx-examples, mlx-lm or mlx if you don’t believe me.
- MLX-LM-LENS: Abliterate and research the inner thoughs of LLMs
- MLX-LM-LoRA: Train models natively on Apple Silicon
- MLX-KAN: KANs implemented natively in MLX
- Local NotebookLM: Run your own "NotebookLM" locally with PDF context
Support for multiple LLM architectures, and training features like:
- OpenBMB's MiniCPM and MiniCPM3,
- Kyutai's Helium,
- State-Space's Mamba v1 and Mamb v2,
- Z.ai & THUKEG's GLM4,
- Rednote dots.llm1,
- Baisu's Ernie4.5 MoE,
- inclusionAI's Bailing MoE and Bailing MoE Linear e.g. Ling-family and Ling-Linear-family,
- IBM's Granite MoE,
- Meituan's LongCat,
- Nvidia's Nemotron H,
- Swiss-AI's Apertus,
- Allenai's OLMoE and Olmo 3,
- Full Weight Fine-Tuning,
- Support for the Muon optimize (both mlx-lm and mlx),
- Reporting training metrics to WandB (Weights & Biases)
- Multiple Optimizers to choose for training
- ReLU² activation function in mlx
- AI21 Lab's Jamba
- And many more ...
A mix of serious deep learning research and fun, experimental tools. Expect anything from multimodal LLMs to dataset labeling UIs and decentralized training methods.
- 🤖 Artificial & General Intelligence
- 🧠 Deep Learning, Transformer Architectures, LoRA magic
- 🎻 Violin — Currently practicing Paganini and Bach
- 🖌️ Occasionally sketching or doing photography for inspiration
- 📺 Huge anime fan (yes, I have seen Berserk)
A Real-Time, Full-Duplex, Multimodal Model that can:
- Control my smart home (sensors, cameras, LEDs, etc.) autonomously
- Understand text, images, videos, and raw audio (like Whisper + GPT-4o but better)
- Output natural speech & engage in open-ended, real-time convos
- Self-update its memory and tools
- Be fully local, private, and offline-first
- Think and act like an autonomous agent with personality
- Autonomous Smart home automation using a Real-Time, Multi-Modal-Large-Language-Action-Model. capable text-audio-vision to text-audio-action (Project J.O.S.I.E.-Home).
- A secret new ML theorem.
- My gabliteration Paper
- Autonomously control and manage my Smart home with sensors, cameras, LEDs, and all the other products.
- Tool calling capabilies.
- Raw audio input with normal user(s) speech and other sounds like (fire, etc.).
- Capable of understanding videos and images (e.g. security cameras).
- Capable of outputting natural assistant audio speech and Text (like GPT4-o).
- Can be Real-Time and Text only like GPT4-o.
- Updating itself Autonomously.
- Having a long-/short-term Memory, and even updating it autonimously.
- Real-Time-full-duplex-Multi-Modal-Model
- Starting a conversation by itself (She donsn't need special user and assistant turns).
- And much more...
- Introduction to Deep Learning & Neural Networks with Keras
- Deep Neural Networks with PyTorch
- Introduction to Computer Vision and Image Processing
- Building Deep Learning Models with TensorFlow
- AI Capstone Project with Deep Learning
- Machine Learning with Python
- Computer Science: Algorithms, Theory, and Machines
- Linear Algebra for Machine Learning and Data Science
- Probability & Statistics for Machine Learning & Data Science
- Unsupervised Learning, Recommenders, Reinforcement Learning
- Advanced Learning Algorithms
- Supervised Machine Learning: Regression and Classification
- CompTIA A+ ... 30 others, and counting.
❤️ From Gökdeniz Gülmez