Hardware
- Mac mini
- The daily inference box. Runs the MLX models and everything wired to provider APIs.
- AI NAS server
- Home for the bigger open weight models and the pipelines that run on a schedule.
Daily drivers
- qwen3.8:27b (MLX)
- Current main driver for research, synthesis and drafting.
- muse-glimmer:30b
- Former driver, kept as the fallback.
On the bench
- deepseek-r1:70b
- All 42GB of it. Candidate for the harness benchmark — heavier reasoning jobs.
Code models
- qwen3-coder:30b
- Main coding model.
- qwen2.5-coder:32b / 14b
- The 14b for quick loops, the 32b when it matters.
Retired, still installed
- qwen3.6 / qwen3.6-hermes
- Superseded as drivers. They stay installed as the control group for every new candidate.
Wired to
- Provider APIs
- Frontier models for the jobs worth paying for, including Gemini for the brand versus competitor search trend watch.
- Quiver + Massive
- Market data and insider trading filings feeding the personal finance analyst.
Questions about any of it: hello@mariocarm0na.com