Instructions to use openvuln/altar-1-4bit-mlx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use openvuln/altar-1-4bit-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("openvuln/altar-1-4bit-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 openvuln/altar-1-4bit-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 "openvuln/altar-1-4bit-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": "openvuln/altar-1-4bit-mlx"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use openvuln/altar-1-4bit-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 "openvuln/altar-1-4bit-mlx"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "openvuln/altar-1-4bit-mlx"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "openvuln/altar-1-4bit-mlx",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use openvuln/altar-1-4bit-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 "openvuln/altar-1-4bit-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 openvuln/altar-1-4bit-mlx

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use openvuln/altar-1-4bit-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 "openvuln/altar-1-4bit-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 "openvuln/altar-1-4bit-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"

Altar-1 (4-bit, MLX)

AikidoSec/altar-1 — the 504B expert-pruned GLM-5.3 (REAP: 168 of 256 routed experts kept, INT4 W4A16 on the experts) — converted to MLX format to run locally on Apple Silicon with stock mlx_lm. No additional quantization was applied: the source INT4 codes and scales are copied bit-identically.

Requires ~330+ GB of unified memory. Developed and tested on a Mac Studio M3 Ultra (512 GB). Expect ~13 tok/s decode at any context length, 126–205 tok/s prefill depending on context length.

Usage

pip install mlx-lm>=0.31
mlx_lm.generate --model openvuln/altar-1-4bit-mlx \
    --prompt "The capital of France is" --max-tokens 32

from mlx_lm import load, generate

model, tokenizer = load("openvuln/altar-1-4bit-mlx")
response = generate(model, tokenizer, prompt="Write a haiku about firewalls", max_tokens=64)
print(response)

OpenAI-compatible server:

mlx_lm.server --model openvuln/altar-1-4bit-mlx
# API on http://127.0.0.1:8080/v1

The custom model file (`glm_moe_dsa_shared.py`)

GLM-5.3's DeepSeek Sparse Attention computes the indexer top-k only on "full" layers; "shared" layers reuse the previous full layer's selection. Materializing per-layer indexer weights (required for a plain checkpoint) changes that behavior and stutters on long contexts. This repo therefore ships glm_moe_dsa_shared.py, which mlx_lm loads automatically via "model_file" in config.json (no trust_remote_code flag needed) to restore true index sharing at run time. Keep the .py file when copying the checkpoint — the model does not load without it.

Fidelity vs AikidoSec/altar-1

INT4 codes, scales, and the computed MLX biases are bit-identical to the source (exhaustive audit, ~112k tensor comparisons).
kv_b_proj is requantized at the same 4 bits / group 32 by mlx-lm's own sanitize() (146 tensors, max deviation 0.856 of one quantization step; mathematical bound 1.0) — identical to official mlx-lm conversions of this family.
The MTP (multi-token prediction) layer that mlx-lm drops is not included.

Performance (Mac Studio M3 Ultra, 512 GB)

context	prefill (tok/s)	TTFT	decode (tok/s)	peak memory
128	86	1.5 s	13.7	315 GB
2k (dense)	205	10.0 s	12.6	319 GB
4k	194	21.1 s	13.0	320 GB
8k (DSA sparse)	167	49.0 s	12.3	323 GB
16k	126	2.2 min	12.6	329 GB
64k	44.5	24.5 min	11.7	367 GB
128k	23.7	92.3 min	11.5	406 GB
256k	~12 (est.)	~6.2 h (est., not run)	~11.5 (est.)	~485 GB (est.)

Serving with mlx_lm.server, decode throughput scales with concurrency (MoE batching): 12.4 tok/s single stream, 18.6 at 2 streams, 26.4 at 4.

License

Inherits the GLM-5.3 license through AikidoSec/altar-1 and cyankiwi/GLM-5.3-AWQ-INT4.

Downloads last month: 111

Safetensors

Model size

495B params

Tensor type

BF16

F32

U32

MLX

Hardware compatibility

4-bit

Model tree for openvuln/altar-1-4bit-mlx

Base model

zai-org/GLM-5.3

Quantized

cyankiwi/GLM-5.3-AWQ-INT4

Quantized

AikidoSec/altar-1

Quantized

(1)

this model

Altar-1 (4-bit, MLX)

Usage

The custom model file (glm_moe_dsa_shared.py)

Fidelity vs AikidoSec/altar-1

Performance (Mac Studio M3 Ultra, 512 GB)

License

Model tree for openvuln/altar-1-4bit-mlx

The custom model file (`glm_moe_dsa_shared.py`)