Instructions to use inferencerlabs/Kimi-K2.6-MLX-Q3i with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use inferencerlabs/Kimi-K2.6-MLX-Q3i 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("inferencerlabs/Kimi-K2.6-MLX-Q3i")
config = load_config("inferencerlabs/Kimi-K2.6-MLX-Q3i")

# 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 inferencerlabs/Kimi-K2.6-MLX-Q3i with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "inferencerlabs/Kimi-K2.6-MLX-Q3i"

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": "inferencerlabs/Kimi-K2.6-MLX-Q3i"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use inferencerlabs/Kimi-K2.6-MLX-Q3i 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 "inferencerlabs/Kimi-K2.6-MLX-Q3i"

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 inferencerlabs/Kimi-K2.6-MLX-Q3i

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use inferencerlabs/Kimi-K2.6-MLX-Q3i with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "inferencerlabs/Kimi-K2.6-MLX-Q3i"

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 "inferencerlabs/Kimi-K2.6-MLX-Q3i" \
  --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"

inferencerlabs commited on Apr 23

Commit

a293bcd

verified ·

1 Parent(s): fe5d6ed

Update README.md

Browse files

Files changed (1) hide show

README.md +6 -5

README.md CHANGED Viewed

@@ -24,17 +24,18 @@ pipeline_tag: image-text-to-text
 <p style="margin-bottom:0px;">
 <strong>Q3.5-INF uses the data-agnostic INF method tuned to yield maximum general accuracy within a 512GB memory budget</strong>
-Due to system memory constraints, the base generation required inferencing directly from SSD and was limited to the first 512 tokens for our coding tests. To address this, generation was configured to output complex functionality from the start rather than incremental scaffolding. While this setup differs from typical usage, it stresses early-token accuracy and the observed trends appear consistent with those seen in larger-scale evaluations of other models. These figures may be updated in future with larger runs.
 </p>
 <table style="border-collapse: collapse; text-align:center; margin-top:10px; margin-bottom:0px;">
 <thead>
 <tr><th>Quantization (bpw)</th><th>Perplexity</th><th>Token Accuracy</th><th>Missed Divergence</th><th>Size</th></tr>
 </thead>
 <tbody>
-<tr><td><strong>Q3.5</strong></td><td>1.1328125</td><td>94.92%</td><td>42.71%</td><td>450.19GB</td></tr>
-<tr><td><strong>Q3.5-INF</strong></td><td>1.078125</td><td>96.67%</td><td>22.04%</td><td>455.68GB</td></tr>
-<tr><td><strong>Q3.6</strong></td><td>1.1484375</td><td>94.72%</td><td>48.72%</td><td>470.99GB</td></tr>
-<tr><td><strong>Base</strong></td><td></td><td>100%</td><td>0.000%</td><td>658.59GB</td></tr>
 </tbody>
 </table>

 <p style="margin-bottom:0px;">
 <strong>Q3.5-INF uses the data-agnostic INF method tuned to yield maximum general accuracy within a 512GB memory budget</strong>
+Due to system memory constraints, the base generation required inferencing directly from SSD and was limited to the first 512 tokens for our coding tests. To address this, generation was configured to output complex functionality from the start rather than incremental scaffolding. While this setup differs from typical usage, it stresses early-token accuracy and the observed trends appear consistent with those seen in larger-scale evaluations of other models. These figures may be updated in future with extended evaluations.
 </p>
 <table style="border-collapse: collapse; text-align:center; margin-top:10px; margin-bottom:0px;">
 <thead>
 <tr><th>Quantization (bpw)</th><th>Perplexity</th><th>Token Accuracy</th><th>Missed Divergence</th><th>Size</th></tr>
 </thead>
 <tbody>
+<tr><td><strong>Q3.5</strong></td><td>1.1328125</td><td>94.92%</td><td>42.71%</td><td>450.19 GB</td></tr>
+<tr><td><strong>Q3.5-INF</strong></td><td>1.078125</td><td>96.67%</td><td>22.04%</td><td>455.68 GB</td></tr>
+<tr><td><strong>Q3.6</strong></td><td>1.1484375</td><td>94.72%</td><td>48.72%</td><td>470.99 GB</td></tr>
+<tr><td><strong>Base</strong></td><td>Untested</td><td>100%</td><td>0.000%</td><td>658.59 GB</td></tr>
 </tbody>
 </table>