Build your model

Make one connected design

This chapter joins the Model Factory decisions into one guided build. You are not ordering hardware or producing a procurement quote. You are creating a coherent design that can be questioned before expensive training begins.

Start with the job. A general assistant needs broad capability. A domain model can be smaller and more focused. An on-device model must fit tight memory, power, and latency limits. Those different purposes should not begin from the same architecture.

Carry one design through the factory

Use the lab to make six connected decisions:

1. Choose the model's purpose.
2. Size its layers, width, vocabulary, and weight precision.
3. Plan a pretraining scale.
4. Choose how people will shape its behavior.
5. Decide how the model will be released.
6. Review the resulting build sheet.

Interactive explanation

Build one model from purpose to artifact

Core + Expert

Model Factory capstone

Build one model from purpose to artifact

interactive explanation

Make six connected decisions. The same selected model moves through architecture, pretraining, alignment, release, and a final build sheet.

01Choose purpose02Size architecture03Plan pretraining04Choose alignment05Decide release06Review build sheet

Current modelDomain specialist

Start with the job. The purpose changes the realistic architecture, precision, and training plan.

General assistantBroad knowledge and conversational ability across many tasks60 layers · 6,656 width · 16-bit target

Capability and evaluation breadth drive scale and cost.

Domain specialistDeep performance in one field with a narrower operating range32 layers · 4,352 width · 8-bit target

High-quality domain data matters more than maximum size.

On-device modelPrivate, low-latency operation on a laptop or phone24 layers · 2,048 width · 4-bit target

Memory, power, and latency impose a hard size ceiling.

The final sheet is the handoff to inference. It identifies the artifact that must be stored, loaded into memory, and served. A model that looks reasonable during training can still be impractical if its weight footprint, runtime support, latency, or release restrictions do not fit the serving system.

Core check

Why choose the model's purpose before selecting its layers and width?

APurpose sets the capability, memory, latency, and cost constraints the design must satisfy.BPurpose automatically trains the final weights.CPurpose removes the need for evaluation.