MLX-LM contributor
Around 10,000 lines of code contributed to Apple’s open-source language model toolkit.
Open-source contributor · Apple MLX
Systems Engineer at Computacenter, specialising in machine learning engineering and MLOps. I build machine-learning systems and help move them from experiments into reliable services. Outside work, I contribute to open source and research ways to run and train AI on Apple chips. I’m a top MLX-LM contributor and, as a freelancer officially sponsored by Neywa Labs, the sole maintainer of MLX-VLM’s core trainer. I also build the J.O.S.I.E. model family and JOSIE-Voice for Mac.
At Computacenter, I work as a System Engineer focused on machine-learning engineering and MLOps—the work of deploying, monitoring and maintaining software that uses AI. Outside work, I contribute to open source and conduct ML research. Apple’s MLX framework is a foundation for building, training and running machine-learning models on Apple chips. MLX-LM adds text language models; MLX-VLM adds models that work with images and words and is now developed at Neywa Labs, which officially sponsors my freelance work on its core trainer. My contributions help those libraries support more models and training methods.
02 / Selected impact
Around 10,000 lines of code contributed to Apple’s open-source language model toolkit.
MLX-LM architecture work spanning more than 15 organizations.
As a freelance engineer officially sponsored by Neywa Labs, I’m the sole maintainer of the core trainer, with SFT, DPO, ORPO and GRPO workflows. MTP and QAT are in open pull requests as the trainer expands to more modalities.
A six-minute MacBook Pro M4 Pro call, shown in the teaser below, with on-device speech models and a cloud reasoning model.
03 / Upstream
MLX is Apple’s open-source toolkit for machine learning. The tools around it help people run and train AI models directly on Apple chips. Here are some of my credited contributions.
I added Muon, a method for updating a model as it learns, and ReLU², a small mathematical function used inside neural networks. Both are credited in MLX’s contributor acknowledgements.
See MLX acknowledgementsMLX-LM is an open-source library in Apple’s MLX ecosystem for running and fine-tuning text-based language models on Apple Silicon. I’m one of its top five contributors, with around 10,000 lines of code added in total. I helped it support more than 20 model designs, and added ways to train every learned value in a model, choose between training methods, and track experiments with Weights & Biases, a tool for recording and comparing training runs. A model design is its internal blueprint; the list below links to the original model files and code changes.
Read MLX-LM creditsMLX-VLM extends the MLX ecosystem to vision-language models that understand images and text together. It now belongs to Neywa Labs, the company building local AI infrastructure for Apple Silicon. As a freelance engineer officially sponsored by Neywa Labs, I develop, create, set up and maintain MLX-VLM’s core trainer, and have added 30,000 lines of code. The trainer supports SFT, DPO, ORPO and GRPO; MTP and QAT are in open pull requests, with more modalities and algorithms planned.
Explore MLX-VLM Visit Neywa Labs See selected trainer PRsMLX Examples offers practical, runnable demonstrations. Its acknowledgements credit my work bringing several models to the examples and adding full training, which adjusts all of a model’s learned values.
Read MLX Examples creditsMLX-LM credits my work on these 25+ model variants from 15+ organizations. An architecture is a model’s internal design; weights are the learned values that let it work. “Weights” links to original model files. “PR” means pull request: a proposed code change and its review.
MiniCPM · MiniCPM3
Weights MiniCPM family · PRs #685 · #24
Qwen3Next
Weights Qwen3 Next · PRs #441 · #453
Helium
Weights Helium-1 preview · 2B · PR #1208
DeepSeek v4.1 Flash
Weights DeepSeek v4.1 Flash · PR #1895
dots.llm1
Weights dots.llm1.inst · PR #211
ERNIE 4.5 MoE
Weights ERNIE-4.5-21B-A3B · PR #267
Bailing MoE (Ling) · Bailing MoE Linear (Ling-Linear)
Weights Ling-lite · Ring Mini Linear · PRs #369 · #513
Klear
Weights Klear-46B-A2.5B · PR #437
Jamba
Weights Jamba-v0.1 · PR #544
Granite MoE
Weights Granite 4.0 H Small · PR #413
Mistral4
Weights Mistral Small 4 · PR #1012
LongCat
Weights LongCat Flash Chat · PR #423
Nemotron H
Weights Nemotron H 8B · PR #407
Apertus
Weights Apertus 8B · PR #421
Lille130m
Weights Lille 130M · PR #429
Qwen3Next · Qwen3 · Qwen3MoE
Weights Qwen3-Next · Qwen3 · PRs #441 · #42* · #199
TeleChat3
Weights TeleChat3 36B · PR #773
Qwen3 note: PR #42 closed before it was merged. The project credits my help on Qwen3 and its version built from specialist submodels; PR #199 is a later accepted update to that version.
Read the MLX-LM acknowledgements04 / Research
A model with several specialist parts that can activate different experts for different inputs. DynaMoE studies whether choosing those specialists dynamically can help a model use its computing resources more efficiently.
Read the paperA method for changing selected model behaviors by editing its learned values, while checking that other abilities still work well.
Read the paperDirectional and Similarity-aware Latent Alignment (DSLA) studies how feedback about better answers can also shape the patterns a model forms internally while processing text.
Read the paper05 / Model family
J.O.S.I.E. is my family of downloadable AI models, exploring reasoning, honesty and independent judgment. J.O.S.I.E.-2 includes models with 2, 4 and 9 billion learned values, trained on Apple chips. Smaller versions, with fewer bits used to store each learned value, need less memory and are easier to run on a personal computer.
Read the J.O.S.I.E.-2 research note Visit the J.O.S.I.E. project page06 / JOSIE
JOSIE-Voice listens and speaks on your Mac, remembers conversations, and can act in Reminders, Calendar, Maps and Messages. The teaser shows a six-minute call on an M4 Pro with more than 100 tool calls: Qwen-3 0.6B handles on-device speech recognition and text-to-speech, while DeepSeek 4.1 Flash runs through Ollama Cloud.
Here is a 6 minute call within JOSIE-Voice on a MacBook Pro M4 Pro using: Qwen-3 0.6B TTS/ASR 4bit and Deepseek 4.1 Flash on ollama cloud. it does 100+ Tool calls with agentic work in the real world. No cuts or speedup's. https://t.co/xlkTtYk3oQ
— Gökdeniz Gülmez (@ActuallyIsaak) October 18, 2026
07 / Selected work
Tools for adapting language models to new tasks on Apple chips. Fine-tuning continues a model’s training with new examples; LoRA does this by learning a small add-on instead of changing every value in the original model.
Explore the projectMLX-VLM runs and trains models that work with images and text on Apple Silicon, and is now part of Neywa Labs. As a freelancer officially sponsored by Neywa Labs, I develop, set up and maintain its core trainer, with SFT, DPO, ORPO and GRPO workflows. MTP and QAT are in open pull requests as the trainer expands to more modalities.
Explore the project Neywa LabsAn open-source Mac app for adapting language models to new tasks, checking their results and managing models on Apple chips.
Explore the projectA personal model family exploring small language models, reasoning and assistant behavior, with releases and research notes on Hugging Face.
Explore J.O.S.I.E.An in-development Mac voice and chat assistant with on-device speech recognition and speech, conversation memory, and actions in everyday apps.
See the teaser Explore the projectDirectional and Similarity-aware Latent Alignment studies how preference feedback can shape a model’s internal representations as well as its answers.
Read the paper draft Explore DSLAA research tool for looking inside language models: how they represent words, which parts of the input they focus on, and how information moves through them.
Explore the projectA notebook that uses your PDFs as the source for summaries and creates audio versions with language and speech tools you choose.
Explore the projectResearch and tools for finding model edits that change selected behaviors while aiming to keep other abilities intact.
Explore the projectAn implementation of Kolmogorov–Arnold Networks, an alternative model design built from simple, learnable functions, using Apple’s MLX toolkit.
Explore the project08 / Career
System Engineer focused on building machine-learning software and keeping it reliable after deployment. MLOps covers the tools and practices for shipping, monitoring and maintaining AI systems.
In my free time, I contribute to open source and pursue ML research and engineering.
09 / References
10 / Elsewhere
I post updates, new research and project releases on X at @ActuallyIsaak.