A budget is a contract
The Memory Dial states the run-owned memory ceiling that the planner may use for model execution. It is evaluated before weights begin moving. If the model's minimum working set cannot fit, Minirun refuses the run and names the constraint.
The budget is not total system memory. The app preserves device headroom for the operating system, interface, framework allocations, and other work it cannot safely attribute to the model plan.
What the floor contains
Every supported runtime has an irreducible working floor: the widest or otherwise required model region, temporary compute state, and model-specific pools needed to make progress. Read-ahead and resident tiers are added only after the floor is funded.
- Working floor
- Memory that cannot be removed while this execution path is active.
- Read-ahead reserve
- Bounded space for regions staged before their consumer reaches them.
- Pinned regions
- Selected model regions kept resident to avoid repeated storage reads.
- Model-specific state
- For example, a routed-expert cache or recurrent state when the adapter supports it.
- Unused headroom
- Budget that remains intentionally unassigned rather than forcing an unsafe plan.
More memory is model-specific
Increasing the dial does not produce one universal effect. A sequential model may pin selected layers; a routed model may retain experts; another architecture may need the extra space for a larger safe staging window. The plan shown in the app describes the selected model rather than applying a generic speed label.
Likewise, lowering the budget can increase bytes read per token or reduce overlap, but only a run on the current device can measure the result. The dial exposes a resource choice; it does not promise a fixed performance curve.
Plan lifecycle
The user selects a value inside the current device limit.
The model adapter assigns the floor, staging, residency, and cache state.
The runner checks the plan against the verified artifact and effective settings.
Memory ownership follows the accepted plan through the run.
Stop, failure, or completion releases run-owned regions before idle.
Choosing a starting point
The three presets are positions on the model's pin ladder, not moods. The planner ranks the units it could hold resident by how many bytes each one saves per resident byte, and the budgets at which one more unit becomes affordable are that ladder's snap points. A preset lands on such a point, so it lands where the plan actually changes.
Floor is the minimum admitted envelope and pins nothing: every layer streams, every token. Balanced is the largest snap point at or below three-fifths of the device ceiling — for DeepSeek V4 on a Mac that can reach it, that is the whole ladder, all 43 deterministic layers and then the output head, which is the last rung there is. Generous takes the highest boundary the plan can still fund, which for V4 is the device ceiling itself, because above the last rung there is nothing left to buy.
These are resource choices rather than quality levels, and that is a measured statement here: forty tokens of V4 logits carried the same digest whether every deterministic layer was resident or every layer was read from the drive. The dial changes what a run costs, not what it says.
A preset above the device ceiling is shown disabled with its deficit rather than hidden, and it never quietly collapses onto the floor — telling you a device can do something it cannot is worse than showing the shortfall. A budget below the floor is refused by name, never clamped upward.