Post

I've been experimenting with local AI (oMLX) on a Mac Studio (M2 Max 96GB) for agentic GitOps. My test bench covers repo analysis and live cluster operations through the Flux MCP server.

Stefan Prodan

@stefanprodan.com

After benchmarking a range of models for a couple of days, I chose Apodex-1.1-mini. I quantized it with Lightning MTP speculative decoding enabled, which made it noticeably faster on my homelab. If you want to try Apodex on oMLX, my quant is on Hugging Face: huggingface.co/stefanprodan...

huggingface.costefanprodan/Apodex-1.1-mini-oQ4e-mtp · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.

1/9/26, 8:02 am0 reposts0 likes

Stefan Prodan

@stefanprodan.com

1 Sept

Some great things came out of this experiment. The Flux MCP we just released ships with major optimizations and new tools for diffing repo changes against live clusters. Check out the new MCP tools here: fluxoperator.dev/docs/mcp/too...

fluxoperator.devTools reference - Flux Operator DocsFlux MCP Server available tools, functions and commands for AI agents

Stefan Prodan

@stefanprodan.com

1 Sept

After Fable and GPT-Sol, the local models felt excruciatingly slow. By next year, I expect the coming wave of lightweight OSS models to shine at agentic GitOps, from manifest generation to handling promotion pipelines across a cluster fleet.