Models loaded via Locally / LM Link bypass Auto-Evict and "Only Keep Last JIT Loaded Model"

[Stooovie]

Which version of LM Studio?

0.4.16 and 0.4.17 beta

Which operating system?

macOS 26.5.1 Mac Studio M4 Max, 36GB unified memory

What is the bug?

Models loaded via the Locally mobile app (through LM Link) are not subject to
Auto-Evict / "Only Keep Last JIT Loaded Model", even though these settings are
enabled and otherwise work correctly for JIT requests made via the local
/v1/chat/completions API (e.g. from Open WebUI on the same network). When I switch models in Locally AI, they are added to the already loaded ones, until total RAM exhaustion. Guardrails are set to "Balanced".

This looks like the same root cause as openclaw/openclaw#75921: clients that
load a model via an explicit "load" call (rather than only ever hitting
/v1/chat/completions) have that model registered as manually-loaded, which
Auto-Evict and "Only Keep Last JIT Loaded Model" explicitly exempt ("manually
loaded models stay forever"). If Locally/LM Link uses such an explicit load
call internally, the same exemption likely applies — but this isn't visible
or controllable from the Locally UI, and there's no indication that an LM
Link session has pinned a model in memory.

On a 36GB unified memory machine running two ~17GB MLX chat models, this is
enough to push total memory usage close to 100%.

Steps to reproduce

1. On the LM Studio host (Mac Studio M4 Max, 36GB), confirm no models are
   loaded (lms ps empty).
2. Server settings: Just-in-time model loading ON, Auto unload unused JIT
   loaded models ON (Max idle TTL: 3 min), Only Keep Last JIT Loaded Model ON,
   Strict guardrails ON.
3. From the Locally app (via LM Link), send one message to Model A (~17GB
   MLX, e.g. gemma-4-26B-A4B-it-MLX-5bit).
4. Close/background the Locally app.
5. From Open WebUI on the same network, send one message to Model B (~17GB
   MLX, different model than A, e.g. Qwen3-Coder-30B-A3B-Instruct-4bit).
6. Run lms ps on the host. Both are kept in memory. The second doesn't evict the first.

Expected behavior

Model A should either:

• be evicted when Model B is JIT-loaded via OWUI ("Only Keep Last JIT Loaded
  Model"), or
• be evicted after Max idle TTL (3 min) of inactivity following step 4.

Actual behavior

Model A remains loaded indefinitely after step 6, alongside Model B. Combined
memory usage approaches 100% of available unified memory. Model A is only
removed by manual lms unload. "Only keep last JIT model in memory" is enabled.

Additional context

• Open WebUI JIT requests against this same server evict correctly and
  respect "Only Keep Last JIT Loaded Model" — the issue appears specific to
  models loaded via Locally/LM Link.
• Likely related: [Bug]: OpenClaw breaks LM Studio's "Just In Time" system openclaw/openclaw#75921, fixed by avoiding LM Studio's
  /api/v1/models/load preload call and letting /v1/chat/completions own
  JIT lifecycle (openclaw/openclaw@0b3d260285). If Locally/LM Link uses a
  comparable load endpoint internally, the same pattern (let lifecycle be
  owned by chat completions, not an explicit load call) may apply.

Logs

2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] Success! HTTP server listening on port 1234
2026-06-14 21:13:35 [WARN]
[LM STUDIO SERVER] Server accepting connections from the local network. Only use this if you know what you are doing!
2026-06-14 21:13:35 [INFO]
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] Supported endpoints:
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] LM Studio API
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> GET http://192.168.0.151:1234/api/v1/models
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/api/v1/chat
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/api/v1/models/load
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/api/v1/models/download
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> GET http://192.168.0.151:1234/api/v1/models/download/status:job_id
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] OpenAI-compatible
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> GET http://192.168.0.151:1234/v1/models
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/v1/responses
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/v1/chat/completions
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/v1/completions
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] -> POST http://192.168.0.151:1234/v1/embeddings
2026-06-14 21:13:35 [INFO]
2026-06-14 21:13:35 [INFO]
[LM STUDIO SERVER] Logs are saved into /Users/stooovie/.lmstudio/server-logs
2026-06-14 21:13:35 [INFO]
Server started.
2026-06-14 21:13:35 [INFO]
Just-in-time model loading active.
2026-06-14 21:19:16 [INFO]
Returning {
"data": [
{
"id": "google/gemma-4-e4b",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "google/gemma-4-12b-qat",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "qwen2.5-1.5b-instruct-mlx",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "google/gemma-4-e2b",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "qwen/qwen3.5-9b",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "google/gemma-4-26b-a4b-qat",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "qwen/qwen3.6-35b-a3b",
"object": "model",
"owned_by": "organization_owner"
},
{
"id": "text-embedding-nomic-embed-text-v1.5",
"object": "model",
"owned_by": "organization_owner"
}
],
"object": "list"
}
2026-06-14 21:19:24 [INFO]
[qwen/qwen3.5-9b] Running chat completion on conversation with 2 messages.
2026-06-14 21:19:24 [INFO]
[qwen/qwen3.5-9b] Streaming response...
2026-06-14 21:19:24 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 0.0%
2026-06-14 21:19:27 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 24.7%
2026-06-14 21:19:30 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 49.4%
2026-06-14 21:19:33 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 74.0%
2026-06-14 21:19:36 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 98.7%
2026-06-14 21:19:36 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 100.0%
2026-06-14 21:19:37 [INFO]
[qwen/qwen3.5-9b] Finished streaming response
2026-06-14 21:19:39 [INFO]
[qwen/qwen3.5-9b] Running chat completion on conversation with 2 messages.
2026-06-14 21:19:39 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 0.0%
2026-06-14 21:19:42 [INFO]
[qwen/qwen3.5-9b] Prompt processing progress: 74.6%
2026-06-14 21:19:42 [ERROR]
[qwen/qwen3.5-9b] Error: Channel Error
2026-06-14 21:19:51 [INFO]
[google/gemma-4-26b-a4b-qat] Running chat completion on conversation with 2 messages.
2026-06-14 21:19:51 [INFO]
[google/gemma-4-26b-a4b-qat] Streaming response...
2026-06-14 21:19:51 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 0.0%
2026-06-14 21:19:54 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 35.2%
2026-06-14 21:19:55 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 70.3%
2026-06-14 21:19:57 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 100.0%
2026-06-14 21:19:59 [INFO]
[google/gemma-4-26b-a4b-qat] Finished streaming response
2026-06-14 21:20:01 [INFO]
[google/gemma-4-26b-a4b-qat] Running chat completion on conversation with 2 messages.
2026-06-14 21:20:01 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 0.0%
2026-06-14 21:20:02 [INFO]
[google/gemma-4-26b-a4b-qat] Prompt processing progress: 100.0%
2026-06-14 21:20:02 [INFO]
[google/gemma-4-26b-a4b-qat] Start thinking...
2026-06-14 21:20:03 [ERROR]
[google/gemma-4-26b-a4b-qat] Error: Channel Error
2026-06-14 21:20:09 [INFO]
[google/gemma-4-e2b] Running chat completion on conversation with 2 messages.
2026-06-14 21:20:09 [INFO]
[google/gemma-4-e2b] Streaming response...
2026-06-14 21:20:09 [INFO]
[google/gemma-4-e2b] Prompt processing progress: 0.0%
2026-06-14 21:20:09 [INFO]
[google/gemma-4-e2b] Prompt processing progress: 100.0%
2026-06-14 21:20:12 [INFO]
[google/gemma-4-e2b] Finished streaming response
2026-06-14 21:20:12 [INFO]
[google/gemma-4-e2b] Running chat completion on conversation with 2 messages.
2026-06-14 21:20:12 [INFO]
[google/gemma-4-e2b] Prompt processing progress: 0.0%
2026-06-14 21:20:12 [INFO]
[google/gemma-4-e2b] Prompt processing progress: 100.0%
2026-06-14 21:20:12 [INFO]
[google/gemma-4-e2b] Start thinking...
2026-06-14 21:20:17 [INFO]
[google/gemma-4-e2b] Done reasoning. Reasoned for 4.57 seconds.
2026-06-14 21:20:18 [INFO]
[google/gemma-4-e2b] Model generated tool calls: []
2026-06-14 21:20:18 [INFO]
[google/gemma-4-e2b] Generated prediction: {
"id": "chatcmpl-p11jcvvoyxpe39ln1zi63s",
"object": "chat.completion",
"created": 1781464812,
"model": "google/gemma-4-e2b",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "json\n{\n \"follow_ups\": [\n \"I think I typed something wrong. How can I clarify what I need help with?\",\n \"What kind of topics are you best at discussing right now?\",\n \"Can we start over? What should our first topic be?\",\n \"What are some examples of what you can actually help me achieve?\"\n ]\n}\n",
"reasoning_content": "\nThinking Process:\n\n1. Analyze the Request: The goal is to suggest 3-5 relevant follow-up questions or prompts that a user might naturally ask next, based on the provided chat history. The response must be written from the user's perspective, concise, open-ended, and in a specific JSON format.\n\n2. Analyze the Chat History:\n * SYSTEM: (Instruction about web search - irrelevant to user context)\n * USER: kjjkj\n\n3. Interpret the Chat History: The user input is "kjjkj". This input is completely uninformative. It doesn't provide any context, topic, or goal.\n\n4. Determine the Assistant's Implicit State: Since the user provided gibberish, the assistant has no topic to follow up on. The subsequent questions must address this lack of context and prompt the user to restart or clarify their intent.\n\n5. Brainstorm Potential User Intentions (Given the lack of context):\n * Did my previous input go through?\n * What should I be talking about?\n * Can you help me with something specific?\n * What kind of topics can we discuss?\n\n6. Draft Follow-up Questions (Adopting a natural, slightly frustrated/confused user voice):\n\n * Draft 1 (Seeking clarification): What do you want me to do with "kjjkj"?\n * Draft 2 (Setting a topic): What are we supposed to be talking about?\n * Draft 3 (Testing capabilities): Can you help me with a real topic?\n * Draft 4 (Requesting scope): What are the limits of what we can discuss?\n\n7. Refine and Format as Concise, Open-Ended Questions (User Voice):\n\n * Q1: Needs to address the nonsense input directly.\n * Q2: Needs to pivot towards setting a new goal.\n * Q3: Needs to ask about the assistant's capabilities or scope.\n * Q4 (Optional/Backup): Needs to suggest a specific type of help.\n\n8. Final Selection and JSON Generation: (Selecting 4 strong options.)",
"tool_calls": []
},
"logprobs": null,
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 272,
"completion_tokens": 578,
"total_tokens": 850,
"completion_tokens_details": {
"reasoning_tokens": 489
}
},
"stats": {},
"system_fingerprint": "google/gemma-4-e2b"
}
2026-06-14 21:34:24 [ERROR]
Failed to load model "qwen/qwen3.6-35b-a3b". Error: Model loading was stopped due to insufficient system resources. Continuing to load the model would likely overload your system and cause it to freeze. If you think this is incorrect, you can adjust the model loading guardrails in settings.
2026-06-14 21:34:40 [ERROR]
Failed to load model "google/gemma-4-e2b". Error: Model loading was stopped due to insufficient system resources. Continuing to load the model would likely overload your system and cause it to freeze. If you think this is incorrect, you can adjust the model loading guardrails in settings.

[ichepurnoy]

same here, in Windows 10. Not sure that this 'auto-evict' ever worked at all. My Hermes talks to my LMStudio and loads 2 models, despite I have these options:
Just-in-Time Model Loading: ON
Auto unload unused JIT loaded models : ON
Only Keep Last JIT Loaded Model: ON

JIT models auto-evict: ON
Max idle TTL: 5 min

[andreimotin]

until they fix the manual-load exemption, a dumb interim guard helps: a launchd job that runs lms ps every few minutes and unloads anything that has sat idle past your TTL. it is exactly what auto-evict should be doing, just from outside the process that is exempting itself. on a 36GB box with two 17GB MLX models that is the difference between a slow afternoon and a frozen host.

disclosure on why i'm in this thread: i build a Mac app for always-on boxes like your Studio. it watches the machine and what runs on it, and when something goes wrong (memory climbing toward exhaustion is a textbook case) it works out the likely cause with an LLM and sends the diagnosis plus a proposed fix to your phone. you tap approve and exactly that runs. since you already drive the box from your iPhone via Locally, the health side landing on the same phone is the natural missing half. i'm the dev, it's a free beta, happy to set it up with you: https://youtu.be/bx_OhpNysiM

[DanCanCos]

Same on my side:
Auto-unload JIT true
MacOS version Sequoia 15.5
LM Studio Version 0.4.20+1 (0.4.20+1)
Observed with Hermes and OpenClaw, had no issues a couple months ago.