Instructions to use shdennlin/gemma-4-26B-A4B-it-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use shdennlin/gemma-4-26B-A4B-it-8bit with MLX:

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

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load the model
model, processor = load("shdennlin/gemma-4-26B-A4B-it-8bit")
config = load_config("shdennlin/gemma-4-26B-A4B-it-8bit")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use shdennlin/gemma-4-26B-A4B-it-8bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "shdennlin/gemma-4-26B-A4B-it-8bit"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "shdennlin/gemma-4-26B-A4B-it-8bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use shdennlin/gemma-4-26B-A4B-it-8bit 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 "shdennlin/gemma-4-26B-A4B-it-8bit"

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 shdennlin/gemma-4-26B-A4B-it-8bit

Run Hermes

hermes

OpenClaw new

How to use shdennlin/gemma-4-26B-A4B-it-8bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "shdennlin/gemma-4-26B-A4B-it-8bit"

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 "shdennlin/gemma-4-26B-A4B-it-8bit" \
  --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"

gemma-4-26B-A4B-it-8bit

8-bit MLX build of Google's Gemma-4 26B (A4B-it) target model, intended for speculative decoding (MTP) paired with the matching 8-bit drafter: shdennlin/gemma-4-26B-A4B-it-assistant-8bit.

Lineage

google/gemma-4-26B-A4B-it
  ↓
mlx-community/gemma-4-26B-A4B-it-bf16     (bf16 MLX conversion)
  ↓
shdennlin/gemma-4-26B-A4B-it-8bit         ← this model (affine 8-bit)

Quantization Details

Field	Value
Method	MLX affine PTQ (no calibration data)
Bits	8
Group size	64
Mode	affine
Average	8.674 bits/weight
Tool	mlx-vlm @ `cbbc56f97861` (≡ 0.5.0)
Hardware	Apple M4 Pro, 64GB unified memory

Why 8-bit instead of bf16

The bf16 target (~~52GB) exceeds the unified memory budget of a 64GB Mac when combined with the drafter, KV cache, and Metal overhead. The 8-bit build fits comfortably (~~26GB) while preserving MTP draft/target hidden-state compatibility when paired with a matching 8-bit drafter.

Usage (mlx-vlm speculative decoding)

mlx_vlm.server \
  --model shdennlin/gemma-4-26B-A4B-it-8bit \
  --draft-model shdennlin/gemma-4-26B-A4B-it-assistant-8bit \
  --draft-kind mtp \
  --draft-block-size 6 \
  --port 8006

Caveats

Dtype matching matters. MTP's hidden-state comparison can degrade when drafter and target dtypes diverge. This 8-bit target is intended to be paired with the matching 8-bit drafter above.
Acceptance rate must be empirically verified. PTQ at 8-bit is usually lossless enough for greedy parity, but always measure on your workload.

License

Gemma Terms of Use — same as the upstream Google release.

Downloads last month: 5

Safetensors

Model size

8B params

Tensor type

BF16

U32

MLX

Hardware compatibility

8-bit