lms-warm

name	lms-warm
description	Unified LM Studio control on Mac — pull, load (with optional max context), warm prefill cache, unload, fresh-reload (unload-then-load), status, list. Use when the user says "warm up <model>", "preload", "prime the cache", "load <model>", "unload", "reload <model>", "fresh load", "switch to <model>", "download <model>", "what's loaded", "max context".
category	execution
runtimes	["claude"]
pii_safe	true

/lms-warm — One command for all LM Studio control

A single shell command (lms-warm) on all 3 machines (<WORKSTATION_HOST>, <INFERENCE_HOST>, max). On Mac it runs locally; on <WORKSTATION_HOST>/generic it passthrough-SSHs to max.

What "warm" actually means

LM Studio caches the prefix tokens (system prompt + early messages) after the first prefill. Warm = load + send a tiny first request so that prefix cache is built. Subsequent requests with the same prefix skip prefill (the slow 30–90s part on a 35B model with a long system prompt).

The win comes from what you cache, not from warming itself. Pre-warm with the exact system prompt you'll use later, and your real session is fast. Pre-warm with a generic "hello" and you've cached nothing useful.

Subcommands

# DEFAULT — load + warm (use the system prompt you'll actually use later)
lms-warm <model-id> "<system prompt>"
lms-warm <model-id> "$(cat plans/my-plan.md)"           # prime with your real plan
lms-warm <model-id> "<system>" --max-ctx                # load with model's max context window
lms-warm <model-id> "<system>" --keep                   # also background-ping every 8 min

# FRESH — unload everything first, then load + warm (clean slot for the new model)
lms-warm fresh <model-id> "<system prompt>"
lms-warm fresh <model-id> "<system>" --max-ctx

# LOAD only (no warm, no unload of others)
lms-warm load <model-id>
lms-warm load <model-id> --max-ctx

# UNLOAD all currently loaded models
lms-warm unload

# DOWNLOAD a new model from HuggingFace (mixed case, MLX by default)
lms-warm pull lmstudio-community/Qwen3-Coder-Next-MLX-4bit
lms-warm pull mlx-community/Some-Model --gguf

# INSPECT
lms-warm --status      # what's currently loaded (= lms ps)
lms-warm --list        # all downloaded models (= lms ls)
lms-warm --stop        # kill the background keep-warm pinger

Max context

--max-ctx flag loads the model with its full context window from this map (verified 2026-05-10):

Model	Max ctx
mlx-community/Qwen3.6-35B-A3B-bf16	262144
wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8	262144
qwen/qwen3-coder-30b	262144
qwen/qwen3-coder-next	262144
qwen/qwen3-next-80b	262144
nemotron-cascade-2-30b-a3b	262144
nvidia/nemotron-3-nano	262144
gemma-4-31b-it-turboquant-mlx	131072
gemma-4-26b-a4b-it-turboquant-mlx	131072
gemma-4-26b-a4b-it	131072
llama-4-scout-17b-16e-instruct	10485760
zai-org/glm-4.7-flash	202752
glm-4.7-flash-mlx@4bit	202752
openai/gpt-oss-20b	131072
liquid/lfm2-24b-a2b	128000

Cost note: max context = bigger KV cache = more RAM. On a 64 GB Mac with the 35B 8bit model loaded (37 GB), full 262K context can push memory hard. Use --max-ctx when you need long-context tasks; skip it for short interactive sessions.

Aliases (in opencode/PI model picker, NOT in this CLI)

The opencode/PI model picker resolves these short aliases. The lms-warm CLI takes the FULL model id (alias resolution doesn't happen in shell).

Alias	Full model id
qwen	wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8
plan-fp16	mlx-community/Qwen3.6-35B-A3B-bf16
build-oq8	wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8
qwen-next	qwen/qwen3-next-80b
coder	qwen/qwen3-coder-30b
coder-next	qwen/qwen3-coder-next
glm	glm-4.7-flash-mlx@4bit
cascade	nemotron-cascade-2-30b-a3b
nemotron	nvidia/nemotron-3-nano
gemma	gemma-4-26b-a4b-it-turboquant-mlx
scout	llama-4-scout-17b-16e-instruct
gpt-oss	openai/gpt-oss-20b
lfm2	liquid/lfm2-24b-a2b

Behavior when invoked

When this skill is called, parse the user's intent and run the matching subcommand via Bash:

• "warm up X" / "preload X" / "prime cache for X" → lms-warm <full-id> [system]
• "switch to X" / "load X fresh" / "reload X" → lms-warm fresh <full-id> [system]
• "load X" (no warm) → lms-warm load <full-id>
• "unload" / "free memory" → lms-warm unload
• "what's loaded" → lms-warm --status
• "what models do I have" → lms-warm --list
• "download X" / "pull X from HF" → lms-warm pull <hf-path>

If the user asks for "max context", append --max-ctx. If the user asks to "keep it warm" or "stay warm", append --keep.

If the user names a model by alias (qwen, coder-next, qwen-next, etc.), resolve to the full id from the table above before passing to lms-warm.

Related

• /plan — the planning skill recommends an executor; output includes the exact lms-warm command to use
• /cplan — Claude planner that uses /plan skill content as system prompt
• LM Studio + opencode/PI configs are synced across all 3 machines (Mac is the host on <PRIVATE_IP>:1234)