The reading discipline
Every model announcement, blog post, or marketing page can be processed through five questions. If you can answer them, you understand the model in the only way that matters for production use. If you can't, the announcement is hiding something.
1. What activates per token?
Total params and active params. Is it dense (total = active) or MoE (active < total)? This determines memory cost and per-token compute. Always answer this first; it controls every subsequent cost estimate.
2. What does latency look like?
Is this a reasoning model with variable, high latency, or a standard model with predictable latency? If it is reasoning, are budgets configurable? Is thinking time billed?
3. Is it locally serveable?
Can it run on consumer hardware (single 24GB or 48GB card, or a Mac Studio)? What is the minimum VRAM at common quantizations? Are GGUF/MLX/AWQ ports available? If you cannot answer, you are dependent on a managed API.
4. Is the benchmark win real for YOUR use case?
Benchmark contamination, cherry-picked metrics, and evaluation-protocol tweaks are rampant. The strongest signal is your own evaluation on your own task. If the announcement headlines a benchmark you don't care about, that's still useful information.
5. Is the capability due to architecture, training, or inference scaffolding?
Locating the change on the four-axis map tells you which parts will transfer. Architecture changes are rare and slow to copy; training and inference changes can be replicated quickly by competitors. This is the single best predictor of whether a "moat" will hold for six months.
The discipline pays off compounding
The first time you apply the five questions, it feels slow. By the tenth model announcement, you can do it in two minutes flat. By then you can also feel which questions the announcement is dodging — and that is usually the most informative thing in the post.