Deploy MiniCPM5-1B with LM Studio

Desktop GUI + OpenAI-compatible local server. On Apple Silicon ships two runtimes:

Required input

Steps

1. Install LM Studio + complete onboarding

brew install --cask lm-studio
open -a "LM Studio"     # accept EULA + pick model source

⚠️ The first launch MUST be GUI — lms (CLI) refuses with Cannot find LM Studio installation until LM Studio has run interactively at least once.

2A. GGUF runtime path

mkdir -p ~/.lmstudio/models/openbmb/MiniCPM5-1B-GGUF
huggingface-cli download openbmb/MiniCPM5-1B-GGUF MiniCPM5-1B-${QUANT}.gguf \
    --local-dir ~/.lmstudio/models/openbmb/MiniCPM5-1B-GGUF/

LMS="/Applications/LM Studio.app/Contents/Resources/app/.webpack/lms"
"$LMS" server start                              # binds 127.0.0.1:1234
"$LMS" load minicpm5-1b --gpu max --context-length 8192 -y
"$LMS" ps                                        # verify the model is loaded

2B. MLX runtime path (Apple Silicon, recommended on Mac)

The MLX runtime needs an MLX-format checkpoint. The only published MLX repo is openbmb/MiniCPM5-1B-MLX (4-bit affine); there is no separate -bf16 / -4bit variant. Either drop that one in as-is, or convert locally from openbmb/MiniCPM5-1B (see minicpm5-deploy-mlx). Then:

# Option A — use the official pre-quantized repo
huggingface-cli download openbmb/MiniCPM5-1B-MLX \
    --local-dir ~/.lmstudio/models/openbmb/MiniCPM5-1B-MLX

# Option B — drop a locally converted directory in
mkdir -p ~/.lmstudio/models/openbmb/MiniCPM5-1B-MLX-${QUANT}
cp -r ./minicpm5-mlx-${QUANT}/* ~/.lmstudio/models/openbmb/MiniCPM5-1B-MLX-${QUANT}/

"$LMS" server start
"$LMS" load minicpm5-1b-mlx${QUANT:+-${QUANT}} --gpu max -y

3. Validate

curl http://127.0.0.1:1234/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "minicpm5-1b",
        "messages": [{"role":"user","content":"1+1=?"}],
        "temperature": 0.7, "top_p": 0.95, "max_tokens": 64
    }'

Expected: "2" in the reply.

For the MLX runtime, a think prompt produces output split into message.reasoning_content (the <think> block) and message.content (the final answer) automatically — that's an MLX-runtime feature, not a model setting.

Think vs nothink

LM Studio 0.4.13's chat-completion endpoint does not propagate chat_template_kwargs.enable_thinking to the GGUF runtime. Instead:

• Default = think mode for both runtimes.
• For nothink with the GGUF runtime, prepend the closing think block manually:

  "messages": [
    {"role":"user","content":"1+1=?"},
    {"role":"assistant","content":"<think>\n\n</think>\n\n"}
  ]

  and the model continues from there.
• MLX runtime: think/answer are auto-split, you don't need to do anything.

Common pitfalls

• MLX runtime not available: only on Apple Silicon. On Intel Mac / Windows / Linux LM Studio, only the GGUF runtime works.

When NOT to use

• Just want CLI / scripted runs → minicpm5-deploy-ollama is leaner
• Production server → minicpm5-deploy-vllm
• No GUI desired → minicpm5-deploy-llama-cpp (llama-server)

Reference

docs/deployment/lmstudio.md