minimax

name	minimax
description	Control the MiniMax-M2.7 JANGTQ vmlx server on <MAC_HOST> (:8765). Best local coder <USER_NAME>owns (78% SWE-bench Verified) but holds ~60 GB Mac memory continuously. Use to start, stop, status-check, restart, tail logs, or smoke-test the server. Trigger phrases — "/minimax", "start minimax", "stop minimax", "minimax status", "is minimax loaded", "free minimax memory".
category	execution
runtimes	["claude"]
pii_safe	true

/minimax — JANGTQ MiniMax-M2.7 server control

MiniMax-M2.7 in JANGTQ format is the best local coding model <USER_NAME>owns — 78% SWE-bench Verified, beats Qwen3.6-27B and 35B-A3B. Runs via vmlx-engine (custom MLX TurboQuant runtime) since LM Studio's stock loader can't handle JANGTQ.

Server lives independently of LM Studio:

	LM Studio	vmlx (this skill)
Port	1234	8765
Models	qwen, glm, gpt-oss, lfm2, scout, gemma	minimax-m2.7-jangtq
Auto-evict	yes (single-model JIT)	no — stays loaded
Memory	varies per loaded model	~60 GB always

Commands

minimax start     # bring server up (~60 GB Mac memory)
minimax stop      # shut down + free memory
minimax restart   # bounce
minimax status    # running? memory? ready?
minimax logs      # tail server log
minimax test      # send a sanity chat request

Script: ~/.local/bin/minimax — deployed on <WORKSTATION_HOST>, generic, and max. SSH-aware (runs launchctl locally on max, via SSH from <WORKSTATION_HOST>/generic).

When to stop it

Stop minimax when you need to load an LM Studio model larger than ~50 GB alongside it. With 128 GB total Mac RAM:

LM Studio model	Size	Coexists with minimax?
qwen-turboquant, glm, gemma, gpt-oss, lfm2, qwen27	≤37 GB	✓
llama-4-scout	61 GB	✗ — minimax stop first

The vmlx server holds ~60 GB Metal-wired memory continuously (per launchd plist KeepAlive: true). LM Studio's unloadPreviousJITModelOnLoad only manages its own port-1234 models; it does NOT touch vmlx.

Endpoints

http://<PRIVATE_IP>:8765/v1/...

• GET /v1/models — list (returns minimax-m2.7-jangtq + full HF id)
• POST /v1/chat/completions — OpenAI format
• POST /v1/messages — Anthropic format (use with Claude Code via ANTHROPIC_BASE_URL)
• POST /v1/embeddings

Use it from clients

Client	How
Claude Code	claude-lms-minimax (sets ANTHROPIC_BASE_URL + --model minimax-m2.7-jangtq)
OpenCode	provider vmlx-jangtq-mac/minimax (or /minimax-m2.7-jangtq)
PI agent	provider vmlx-jangtq-mac/minimax-m2.7-jangtq
curl	examples above

Persistence

Managed by launchd plist ~/Library/LaunchAgents/com.<ORG_NAME>.vmlx-jangtq.plist on max. RunAtLoad: true, KeepAlive: true — survives reboots, auto-restarts on crash. Logs at ~/.local/share/jangtq/logs/vmlx.log.

Sampling defaults baked in

vmlx server was started with these defaults (per MiniMax recommendations):

• --default-temperature 1.0
• --default-top-p 0.95
• --reasoning-parser auto (resolves to qwen3 for MiniMax)
• --enable-auto-tool-choice (tool calling enabled with minimax parser)

Per-request overrides work normally — these are just the floor when client doesn't specify.

Troubleshooting

Server not responding → minimax restart. If still dead, minimax logs for the actual error.

Out of memory at load → check vm_stat on Mac. Probably an LM Studio model is loaded — unload it (ssh <MAC_HOST> '~/.lmstudio/bin/lms unload --all') then minimax restart.

Slow first response → first request after load is ~6-8s (Metal kernel JIT cache cold). Subsequent requests fast. Acceptable.

Want to change served port / model → edit ~/Library/LaunchAgents/com.<ORG_NAME>.vmlx-jangtq.plist then minimax restart.