Instructions to use Regis-RCR/gemma-4-31B-it-oQ8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Regis-RCR/gemma-4-31B-it-oQ8 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("Regis-RCR/gemma-4-31B-it-oQ8")
config = load_config("Regis-RCR/gemma-4-31B-it-oQ8")

# 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 Regis-RCR/gemma-4-31B-it-oQ8 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Regis-RCR/gemma-4-31B-it-oQ8"

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": "Regis-RCR/gemma-4-31B-it-oQ8"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use Regis-RCR/gemma-4-31B-it-oQ8 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 "Regis-RCR/gemma-4-31B-it-oQ8"

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 Regis-RCR/gemma-4-31B-it-oQ8

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Regis-RCR/gemma-4-31B-it-oQ8 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Regis-RCR/gemma-4-31B-it-oQ8"

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 "Regis-RCR/gemma-4-31B-it-oQ8" \
  --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"

Regis-RCR commited on Apr 17

Commit

f3ecdad

verified ·

1 Parent(s): 5b1b36e

Update README.md

Browse files

Files changed (1) hide show

README.md +166 -2

README.md CHANGED Viewed

@@ -10,16 +10,180 @@ tags:
 - MLX
 ---
 <p>
-This model is a straight MLX oQ8 quant of google/gemma-4-31B-it for local Apple Silicon inference.
 This model was quantized using (<a href="https://github.com/jundot/omlx" target="_blank">oMLX</a> v0.3.5) mixed-precision quantization.
 No other edits, additions, merges, or behavioral changes have been made to the model beyond the quantization/export step.
 </p>
 <div align="center">
   <img src=https://ai.google.dev/gemma/images/gemma4_banner.png>
 </div>

 - MLX
 ---
+# Gemma Quantized Models Benchmark Card
 <p>
+This repository provides benchmark results for several quantized Gemma instruction-tuned variants evaluated on reasoning, truthfulness, coding, and live coding tasks.
 This model was quantized using (<a href="https://github.com/jundot/omlx" target="_blank">oMLX</a> v0.3.5) mixed-precision quantization.
 No other edits, additions, merges, or behavioral changes have been made to the model beyond the quantization/export step.
 </p>
+## Benchmark Summary
+<div style="overflow-x:auto;">
+  <table>
+    <thead>
+      <tr>
+        <th align="left">Benchmark</th>
+        <th align="left">Mode</th>
+        <th align="left">Sampled</th>
+        <th align="right">E2B oQ8</th>
+        <th align="right">E4B oQ8</th>
+        <th align="right">26B-A4B oQ8</th>
+        <th align="right">31B oQ8</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr>
+        <td><strong>MMLU</strong></td>
+        <td>Sample</td>
+        <td>100 / 14042</td>
+        <td align="right">55.0%</td>
+        <td align="right">71.0%</td>
+        <td align="right">85.0%</td>
+        <td align="right">87.0%</td>
+      </tr>
+      <tr>
+        <td><strong>TRUTHFULQA</strong></td>
+        <td>Full</td>
+        <td>817</td>
+        <td align="right">60.0%</td>
+        <td align="right">72.5%</td>
+        <td align="right">85.4%</td>
+        <td align="right">89.5%</td>
+      </tr>
+      <tr>
+        <td><strong>HUMANEVAL</strong></td>
+        <td>Full</td>
+        <td>164</td>
+        <td align="right">80.5%</td>
+        <td align="right">89.6%</td>
+        <td align="right">96.3%</td>
+        <td align="right">95.1%</td>
+      </tr>
+      <tr>
+        <td><strong>LIVECODEBENCH</strong></td>
+        <td>Sample</td>
+        <td>30 / 1055</td>
+        <td align="right">33.3%</td>
+        <td align="right">50.0%</td>
+        <td align="right">50.0%</td>
+        <td align="right">73.3%</td>
+      </tr>
+    </tbody>
+  </table>
+</div>
+## Per-model Details
+<details>
+  <summary><strong>gemma-4-E2B-it-oQ8</strong></summary>
+  <br>
+  <table>
+    <thead>
+      <tr>
+        <th align="left">Benchmark</th>
+        <th align="right">Accuracy</th>
+        <th align="right">Correct</th>
+        <th align="right">Total</th>
+        <th align="right">Time (s)</th>
+        <th align="left">Think</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr><td>MMLU</td><td align="right">55.0%</td><td align="right">55</td><td align="right">100</td><td align="right">42.9</td><td>No</td></tr>
+      <tr><td>TRUTHFULQA</td><td align="right">60.0%</td><td align="right">490</td><td align="right">817</td><td align="right">293.9</td><td>No</td></tr>
+      <tr><td>HUMANEVAL</td><td align="right">80.5%</td><td align="right">132</td><td align="right">164</td><td align="right">302.9</td><td>No</td></tr>
+      <tr><td>LIVECODEBENCH</td><td align="right">33.3%</td><td align="right">10</td><td align="right">30</td><td align="right">165.4</td><td>No</td></tr>
+    </tbody>
+  </table>
+</details>
+<details>
+  <summary><strong>gemma-4-E4B-it-oQ8</strong></summary>
+  <br>
+  <table>
+    <thead>
+      <tr>
+        <th align="left">Benchmark</th>
+        <th align="right">Accuracy</th>
+        <th align="right">Correct</th>
+        <th align="right">Total</th>
+        <th align="right">Time (s)</th>
+        <th align="left">Think</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr><td>MMLU</td><td align="right">71.0%</td><td align="right">71</td><td align="right">100</td><td align="right">59.6</td><td>No</td></tr>
+      <tr><td>TRUTHFULQA</td><td align="right">72.5%</td><td align="right">592</td><td align="right">817</td><td align="right">336.3</td><td>No</td></tr>
+      <tr><td>HUMANEVAL</td><td align="right">89.6%</td><td align="right">147</td><td align="right">164</td><td align="right">422.5</td><td>No</td></tr>
+      <tr><td>LIVECODEBENCH</td><td align="right">50.0%</td><td align="right">15</td><td align="right">30</td><td align="right">247.3</td><td>No</td></tr>
+    </tbody>
+  </table>
+</details>
+<details>
+  <summary><strong>gemma-4-26B-A4B-it-oQ8</strong></summary>
+  <br>
+  <table>
+    <thead>
+      <tr>
+        <th align="left">Benchmark</th>
+        <th align="right">Accuracy</th>
+        <th align="right">Correct</th>
+        <th align="right">Total</th>
+        <th align="right">Time (s)</th>
+        <th align="left">Think</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr><td>MMLU</td><td align="right">85.0%</td><td align="right">85</td><td align="right">100</td><td align="right">75.1</td><td>No</td></tr>
+      <tr><td>TRUTHFULQA</td><td align="right">85.4%</td><td align="right">698</td><td align="right">817</td><td align="right">404.9</td><td>No</td></tr>
+      <tr><td>HUMANEVAL</td><td align="right">96.3%</td><td align="right">158</td><td align="right">164</td><td align="right">340.6</td><td>No</td></tr>
+      <tr><td>LIVECODEBENCH</td><td align="right">50.0%</td><td align="right">15</td><td align="right">30</td><td align="right">421.7</td><td>No</td></tr>
+    </tbody>
+  </table>
+</details>
+<details>
+  <summary><strong>gemma-4-31B-it-oQ8</strong></summary>
+  <br>
+  <table>
+    <thead>
+      <tr>
+        <th align="left">Benchmark</th>
+        <th align="right">Accuracy</th>
+        <th align="right">Correct</th>
+        <th align="right">Total</th>
+        <th align="right">Time (s)</th>
+        <th align="left">Think</th>
+      </tr>
+    </thead>
+    <tbody>
+      <tr><td>MMLU</td><td align="right">87.0%</td><td align="right">87</td><td align="right">100</td><td align="right">306.4</td><td>No</td></tr>
+      <tr><td>TRUTHFULQA</td><td align="right">89.5%</td><td align="right">731</td><td align="right">817</td><td align="right">846.1</td><td>No</td></tr>
+      <tr><td>HUMANEVAL</td><td align="right">95.1%</td><td align="right">156</td><td align="right">164</td><td align="right">1072.2</td><td>No</td></tr>
+      <tr><td>LIVECODEBENCH</td><td align="right">73.3%</td><td align="right">22</td><td align="right">30</td><td align="right">968.6</td><td>No</td></tr>
+    </tbody>
+  </table>
+</details>
+## Evaluation Notes
+- MMLU and LiveCodeBench were run on sampled subsets rather than the full benchmark.
+- TruthfulQA and HumanEval were evaluated on the full set shown in the table above.
 <div align="center">
   <img src=https://ai.google.dev/gemma/images/gemma4_banner.png>
 </div>