Run it on your Mac.
Five steps, no cloud.

1. 1

   Install uv and get the code

   uv manages Python and pins every package to the versions this build was tested with (MLX 0.32.4, mlx-vlm 0.7.9). The repo holds the server, the memory cap and the dashboard.

   curl -LsSf https://astral.sh/uv/install.sh | sh
   git clone https://github.com/Sauhard74/glm-5.3-flash-sauhard.git
   cd glm-5.3-flash-sauhard
   uv sync

2. 2

   Download the weights

   71 GB from Hugging Face. The download resumes if it is interrupted; run the same command again.

   uv run hf download Sauhard74/GLM-5.3-Flash-sauhard-MLX --local-dir models/glm-5.3-flash-sauhard

3. 3

   Start the server

   Serves the OpenAI API (/v1/chat/completions) and the Anthropic API (/v1/messages) on port 8080, capped at 84 GB. Loading takes a minute or two; the vision warning in the log is expected, since this build is text-only.

   # a key so only your apps can use the server
   openssl rand -hex 24 > control/api_key
   CAP_GB=84 MTP=0 scripts/quickserve.sh models/glm-5.3-flash-sauhard

4. 4

   Say hello

   In a second terminal. The model field must be the same path the server was started with.

   curl -s http://127.0.0.1:8080/v1/chat/completions \
     -H "Authorization: Bearer $(cat control/api_key)" -H "Content-Type: application/json" \
     -d '{"model":"models/glm-5.3-flash-sauhard","messages":[{"role":"user","content":"Write a haiku about unified memory."}],"max_tokens":800}'

5. 5

   Connect Claude Code

   Runs Claude Code against the local model, with a 128K context window. Subagents use it too.

   MODEL_ID=models/glm-5.3-flash-sauhard scripts/claude-local.sh

Recommended sampling: temperature 1.0, top-p 0.95. The model thinks before it answers, so allow at least 2,000 output tokens.

scripts/install-control.sh             # open http://127.0.0.1:8090, pick GLM 5.3 Flash-sauhard, Start
scripts/install-control.sh --uninstall # remove it

The dashboard's second profile, Qwen3.8-Flash-Next, needs its own download; see control/profiles.json.

The server is killed while loading, or the Mac swaps heavily

Close large apps (browsers with many tabs, Docker, other models) before starting. The model needs about 72 GB resident; macOS kills a process that pushes memory pressure too high.

"model not found" or the model loads twice

The model field in requests must match the path passed to quickserve.sh exactly.

Slow first reply

The first request after loading warms the Metal kernels. Long prompts prefill at about 630 tokens per second, so a 30K-token codebase takes about a minute; repeated prefixes are cached.

401 from the server

Send the key as Authorization: Bearer … (OpenAI style) or x-api-key (Anthropic style).

# 1. the original checkpoint
uv run hf download zai-org/GLM-5.3-Flash --local-dir models/src-glm-5.3-flash
# 2. calibration set: public code, agent, math and long-document data
uv run python -m calib_data.build --model glm --out artifacts/glm/data
# 3. stream the model layer by layer, recording routing and sensitivity (~5.5 h)
uv run python -m calib.run --model glm --src models/src-glm-5.3-flash --data artifacts/glm/data --out artifacts/glm/calib
# 4. prune to 153 experts, allocate bits to a 76 GB budget, LSQ rounding (~10 min)
uv run python -m build.make --model glm --src models/src-glm-5.3-flash --calib artifacts/glm/calib \
  --prune 153 --score reap --budget-gb 76 --drop-vision --rounding lsq --out models/glm-5.3-flash-sauhard
# 5. check it against the original
uv run python -m evals.kl --model models/glm-5.3-flash-sauhard

Change --budget-gb for a smaller or larger build, or edit calib_data/recipes.py to calibrate on a different mix of work.