ai.georgeliu.com/p/running-google-gemma-4-locally-with

LM Studio's official 0.4.0 release post says that version introduced the new headless llmster daemon, parallel requests, and a stateful /v1/chat endpoint. It does not list an Anthropic-compatible /v1/messages endpoint as part of 0.4.0.

LM Studio's official API changelog then lists 0.4.1 as the release that added the Anthropic-compatible API: POST /v1/messages. LM Studio's separate Claude Code announcement says the same thing explicitly: "With LM Studio 0.4.1, we're introducing an Anthropic-compatible /v1/messages endpoint."

So the article is off by one release: the Anthropic-compatible endpoint existed by the time the article was published, but it was added in 0.4.1, not 0.4.0.

LM Studio's TTL docs distinguish between JIT-loaded models and models loaded manually with lms load.

• For JIT-loaded models, the docs say the default TTL is 60 minutes.
• But for models loaded with lms load, the docs say they do not have a TTL by default and stay in memory until manually unloaded, unless you explicitly pass --ttl.

So in the context of lms load google/gemma-4-26b-a4b --ttl 1800, saying "The default is 3600 seconds (1 hour)" is incorrect. A 1-hour TTL is something you can set manually with --ttl 3600; it is not the default behavior for lms load.