Model Family — AI Glossary

Reference details

Related terms

• Model Version
• Model Snapshot
• Model Architecture

How it works

1. A family shares a pretraining recipe: the same tokenizer, the same architecture class, and usually the same or an overlapping data mixture. Size tiers are produced by scaling width and depth, not by changing the approach.
2. The tokenizer is the hard constraint. Shared vocabulary means a prompt tokenises identically across tiers, so token counts and therefore cost estimates transfer between members; a family that changed tokenizers mid-generation breaks that assumption.
3. Post-training is applied per tier but from a common specification — the same instruction format, the same system-prompt handling, the same tool-calling schema — which is what makes the API surface interchangeable.
4. Tiers are positioned by a capability-versus-cost trade: a small tier is distilled from or trained alongside a large one, so it inherits behavioural conventions while losing headroom on hard reasoning.
5. Generations version the recipe. A new generation may keep the family name while changing architecture, context length or tokenizer, which is why the family name alone is not a compatibility guarantee.
6. Snapshots pin a specific trained artefact inside a tier, so 'the family' names a lineage, 'the tier' names a size, and only the snapshot names what will actually answer your request.

Example: What swapping tiers actually buys, and what it does not

A pipeline runs 100,000 requests a month at 2,000 input and 500 output tokens each. Suppose the large tier costs 15 units per million input tokens and 75 per million output, and the small tier 1 and 5. Input is 100,000 x 2,000 = 200M tokens; output is 100,000 x 500 = 50M. The large tier costs 200 x 15 + 50 x 75 = 3,000 + 3,750 = 6,750 units. The small tier costs 200 x 1 + 50 x 5 = 200 + 250 = 450 units — a 15x reduction for a one-line model-string change, because the prompt, the tool schemas and the parsing all transfer. What does not transfer is accuracy: if the small tier resolves 88% of requests against the large tier's 96%, then 100,000 x 0.08 = 8,000 requests a month need an escalation path, and the cost of building and operating that path is the real comparison, not the 6,300-unit gap.

Commonly confused with

A family is not a version and not a snapshot, and the three are routinely collapsed in production configs. The family is the lineage, the tier is the size position within it, the version is the generation of the recipe, and the snapshot is the immutable artefact. 'We use the Claude family' says almost nothing operationally; a pinned snapshot says exactly what will run. Teams that pin only the family or tier discover the distinction when a generation rolls forward and prompt-sensitive behaviour shifts underneath them.

When to use it

Reach for it when:

• Planning a cost-versus-capability ladder, where cheap tiers handle volume and expensive tiers handle escalations behind one integration.
• Estimating token cost across tiers, which is only sound when the tokenizer is shared.
• Choosing a vendor for a long-lived integration: a family with consistent conventions across tiers reduces the cost of changing your mind later.
• Reasoning about a provider's roadmap, since families telegraph which capabilities are being invested in.

Reach for something else when:

• As a substitute for naming the snapshot in production. Family-level configuration is how silent behaviour changes reach users.
• As evidence of capability. Family membership predicts API shape, not benchmark performance; a small tier is not a slightly worse large tier on every axis.
• Across generations without re-testing. A new generation under the same family name may change context length, tokenizer or tool-calling format.
• When comparing across vendors. Tier names — mini, small, flash, haiku — are marketing positions, not a shared scale, so cross-family tier comparison is not meaningful.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Model Family — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path