Gökdeniz Gülmez

At Computacenter, Gülmez focuses on deployment, monitoring and maintenance of machine-learning systems. His independent work contributes model architectures and training methods to MLX libraries. Neywa Labs sponsors his freelance maintenance of MLX-VLM’s core trainer. His professional background also includes vocational training in IT systems integration and studies at IT.Schule Stuttgart.[1][4]

Test whether a model understands the MLX framework

MLX Benchmark evaluates language models on questions, coding and debugging tasks about MLX. A command-line runner selects a model and provider, filters tasks by category and difficulty, and stores results for tables and charts. Its judge model can be configured separately from the candidate; the default uses the candidate model itself. The resulting scores therefore need to be interpreted with the task set, grading setup and sample outputs in view.[5]

Study what post-training changes in a small model

His JOSIE-2 report describes carefully selected post-training data for compact language models, including reasoning examples and instruction-following behavior. It compares reasoning and non-reasoning modes and discusses changes that transfer beyond the explicit format of a reasoning trace. In matched-prompt tests, he also reports longer reasoning traces after quantization, leaving the cause unresolved. The release treats model behavior and evaluation as part of a local-first development process.[6]