Qwen 3.5 9B
RecommendedStrongest general local profile across repository, vision, sports, and grounded tasks.
Repeated-trial coverage
39 / 39Apple Silicon · 24 GiB · localhost only
Measure local models where they actually run.
A reproducible MLX lab for comparing one model at a time—complete with process safety, OpenAI-compatible APIs, realistic benchmarks, and evidence you can inspect.
One model at a timeComparable runs without hidden contention
127.0.0.1 by designInference stays on the machine
Raw evidence retainedReports never replace measurements
Current lab snapshot
Different strengths. Same hardware.
The useful question is not “which model wins?” It is which measured behavior fits the workload—and whether the evidence is complete enough to trust the answer.
Gemma 4 12B
MeasuredExcellent reasoning results, with weaker structured-schema reliability in this setup.
Reasoning route score
100%GPT-OSS 20B
IncompleteFast and capable in observed text runs, but repeated-trial coverage remains unfinished.
Repeated-trial coverage
13 / 39