Spark-X2.6-4B launches with 1M context

The compelling story is deployment efficiency: a small open model that can handle daily agent work locally, while leaving heavier workloads for cloud inference. But 1M context is a capacity ceiling, not proof of cheap or reliable million-token reasoning.

• –The architecture is unchanged from Spark-X2.5; reported gains come from post-training and none/low/high thinking controls.
• –Apache 2.0 licensing and broad runtime support make experimentation and self-hosting straightforward.
• –Vendor-reported evaluations claim a 63.3 average, ahead of Spark-X2.5-4B and Qwen3.5-9B, but independent testing is still limited.
• –Long-context quality appears to degrade substantially at 1M tokens, so retrieval, context compression, and routing still matter.
• –Cloud bursting is an orchestration pattern around the model, not a bundled service or guaranteed pennies-per-task feature.