Instructions to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("Sauhard74/GLM-5.3-Flash-sauhard-MLX")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Sauhard74/GLM-5.3-Flash-sauhard-MLX"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Sauhard74/GLM-5.3-Flash-sauhard-MLX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "Sauhard74/GLM-5.3-Flash-sauhard-MLX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "Sauhard74/GLM-5.3-Flash-sauhard-MLX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "Sauhard74/GLM-5.3-Flash-sauhard-MLX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Sauhard74/GLM-5.3-Flash-sauhard-MLX"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Sauhard74/GLM-5.3-Flash-sauhard-MLX

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Sauhard74/GLM-5.3-Flash-sauhard-MLX with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Sauhard74/GLM-5.3-Flash-sauhard-MLX"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Sauhard74/GLM-5.3-Flash-sauhard-MLX" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

GLM 5.3 Flash-sauhard (MLX)

GLM-5.3-Flash (321B mixture-of-experts) rebuilt to run on a 128 GB Apple Silicon Mac with room left for your editor, builds and Claude Code. 71.6 GB resident while serving.

Pruned: 153 of 288 routed experts per layer, chosen by REAP on calibration data from code, agent, math and long-document work.
Quantized: a bit width per tensor from 3 to 8 bits (3.48 bits per weight on average), allocated by measured sensitivity, with LSQ rounding (each scale refit against calibration activations).
Text-only: the vision tower is dropped. GLM's MTP draft head is included (1.6 GB) but speculative decoding has not been verified yet.

Write-up: https://glm-flash-sauhard.vercel.app · Benchmarks: https://glm-flash-sauhard.vercel.app/report/ · Setup: https://glm-flash-sauhard.vercel.app/run/ · Code: https://github.com/Sauhard74/glm-5.3-flash-sauhard

Results

Measured on a MacBook Pro M5 Max (128 GB). Thinking on, temperature 1.0, top-p 0.95, one sample per problem. These are subsets, so ranges are wide.

Benchmark	Problems	Score
HumanEval+	40	95.0%
MBPP+	40	92.5%
MATH-500	40	97.5%
BFCL function calling	105	84.8%

Fidelity against full-precision GLM-5.3-Flash on 48 held-out conversations (522K tokens): KL divergence 0.108 and 92.4% top-1 agreement on the tokens the model writes.

Speed and memory: see the benchmark report.

Run it

Needs an Apple Silicon Mac with 128 GB of unified memory and the serving scripts from the code repo (they cap memory and disable a 22 GB long-context prefill buffer that would otherwise push the model past 84 GB).

git clone https://github.com/Sauhard74/glm-5.3-flash-sauhard.git && cd glm-5.3-flash-sauhard
uv sync
uv run hf download Sauhard74/GLM-5.3-Flash-sauhard-MLX --local-dir models/glm-5.3-flash-sauhard
openssl rand -hex 24 > control/api_key
CAP_GB=84 MTP=0 scripts/quickserve.sh models/glm-5.3-flash-sauhard

OpenAI API at http://127.0.0.1:8080/v1, Anthropic API at http://127.0.0.1:8080/v1/messages. The model field is the path above. Claude Code: MODEL_ID=models/glm-5.3-flash-sauhard scripts/claude-local.sh.

Tested with MLX 0.32.4 and mlx-vlm 0.7.9 (model_type: glm5_next).

Limits

Not run on public leaderboards. The original GLM-5.3-Flash scores 42 on the Artificial Analysis index (medium effort); this build is expected to land near 40–41, but that is an estimate.
Long-context retrieval beyond 64K tokens has not been tested.
Calibration data is public Hugging Face datasets only; no private data was used.

License

MIT, as the base model. Built from zai-org/GLM-5.3-Flash.

Downloads last month: 233

Safetensors

Model size

171B params

Tensor type

U32

F32

BF16

MLX

Hardware compatibility

6-bit

Model tree for Sauhard74/GLM-5.3-Flash-sauhard-MLX

Base model

zai-org/GLM-5.3-Flash

Quantized

(210)

this model