LongCat Flash Lite Sparse

LongCat Flash Lite Sparse adds sparse attention and native long-context support to the non-thinking LongCat Lite family.

Sparse attention for extended history

This 69B mixture-of-experts checkpoint activates approximately 3B parameters per token. It replaces the earlier Lite model’s dense MLA path with LongCat Sparse Attention. The publisher describes native context support up to about one million tokens and a progressive long-context training schedule.

Check the released configuration

The supplied configuration sets 983,040 positions. This profile records that exact setting while describing the broader one-million-token positioning in prose. The chosen serving recipe, cache policy and multimodal support should be checked independently rather than copied from another LongCat release.

Long-context evaluation

Give the model a task that depends on facts separated across the supplied history, then introduce a correction late in the conversation. Check whether the final answer reflects that correction and preserves the earlier requirements. For coding, include an instruction that prevents an unrelated change and inspect the actual diff.

Infrastructure and comparison

Sparse attention can change the work required for a long prompt, but it does not remove the need to hold the full weights and runtime state. Measure prefill, generation and complete task time at your expected concurrency. The weights are MIT-licensed. Compare this checkpoint with ordinary Flash Lite using the same workload and runtime options wherever both support them, keeping context extension and parser settings explicit.

Sources

Model documentation