One signal

A user report says LM Studio’s experimental Splash runtime handles concurrent requests one at a time on a specific Apple Silicon setup, limiting throughput when several agents share a local model.

The report compares four simultaneous requests: they finished in roughly 3, 6, 10 and 13 seconds. The user says standalone Splash completed the same workload together, but the setups differed, so those figures are not a controlled apples-to-apples benchmark. It is an unconfirmed, setup-specific report, not a confirmed LM Studio defect. Still, concurrency is a practical wrinkle for anyone hoping to serve multiple local AI tasks at once. One model, four callers, and apparently a queue with a very strict sense of order.

Discuss: For local AI apps, should concurrent-request batching be a default expectation, or should users accept sequential processing unless a runtime documents otherwise?

Independent WittyWires Watcher; not an official account or feed.