Exact context windows

Set the number you mean

Load Lab accepts an exact custom integer. To request 57,000 tokens, enter 57000. Presets are conveniences; they do not restrict you to powers of two. The app preserves the chosen integer through its load request rather than replacing it with the nearest preset.

The Lemonade path accepts 128 through 262,144 tokens. The experimental Strata adapter has a narrower exact range of 256 through 8,192; changing engines retains the input and reports an over-limit error instead of shortening it. The selected artifact must also expose a known context ceiling, and the request cannot exceed it. Blank, fractional, negative and over-limit values are invalid. A model declaring 262,144 is not a promise that your hardware can allocate that context.

What context includes

The context budget accommodates the model’s input representation: instructions, chat history, current user input and space for generated tokens. Actual tokenization and chat templates are model-specific. Characters, words and tokens are different units.

Longer context can increase state memory and prompt-processing time. Hybrid and recurrent architectures complicate the relationship: full-attention KV is only part of their state. An apparently small attention-cache estimate must not be read as total runtime memory.

Exactness has a verification gate

The qualified backend is launched with automatic context fitting disabled. That protects the selected value from an engine silently reducing it to fit. It can also make a load fail rather than quietly succeeding at a shorter window. The truthful outcome is either the requested effective context or a visible load error.

For the tested RegesCore path, runtime evidence included ctx_size=8192 and n_ctx_slot=8192. That proves the 8,192 smoke-test configuration. It does not qualify a 57,000 or 262,144 allocation, nor long-context retrieval quality.

Practical workflow

1. Start with a context appropriate to the task and a known-good load.
2. Set the exact new value and recalculate the memory plan.
3. Unload/reload deliberately and check the effective engine settings.
4. Test a workload near the intended length, including output space.
5. Record first-token latency, state allocation and answer quality separately.

If a conversation becomes too large, start a new chat or reduce its content. Automatic summarization, context compaction and exact token-budget previews are future work. Do not assume an unseen compactor preserves every instruction or fact.