Instructions to use ldenoue/title-onnx with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ldenoue/title-onnx 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("ldenoue/title-onnx")

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 ldenoue/title-onnx with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "ldenoue/title-onnx"

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": "ldenoue/title-onnx"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use ldenoue/title-onnx with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "ldenoue/title-onnx"

Run an OpenAI-compatible server

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

Hermes Agent

How to use ldenoue/title-onnx 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 "ldenoue/title-onnx"

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 ldenoue/title-onnx

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ldenoue/title-onnx with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "ldenoue/title-onnx"

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 "ldenoue/title-onnx" \
  --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"

Title

Suggest a title and description for any text.

On-device titles and descriptions: a short factual title and a one- to two-sentence description for any passage of text.


Platforms	iOS, macOS, tvOS, visionOS
Weights	v0.1.0

Get a title and a one or two sentence description for any passage of text, on device.

Title is fine-tuned on transcript clips, but it works on any prose. The register is deliberately plain, with no emoji, no hashtags and no clickbait, and a description identifies this passage rather than its topic.

Install

Swift (requirements)

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "4.2.3",
        traits: ["MLX"])

Then add the Title product to your target. The MLX trait is required: without it the module compiles as a stub.

Usage

Title runs only on Apple platforms, on the GPU through MLX. The other Desert Ant models on Apple platforms run on Core ML. We measured Title writing text 5.7-8.3x faster on the GPU than on the Neural Engine. Because Title needs MLX, you turn Title on with the MLX package trait. A build that leaves out the trait fails at compile time, so you can't ship an app without a working model by mistake.

Loading the model is slow and writing a card is fast, so create one Titles and reuse it. Titles is an actor, because MLX isn't safe to drive from several tasks at once. Calls from several tasks run one at a time.

On iOS, call suspend() before your app enters the background and resume() when it returns. iOS revokes GPU access in the background, and MLX crashes the app if a generation is still running. suspend() cancels the running generation, and the call retries after resume().

Swift

Title doesn't download its weights. Download the model files from Hugging Face, and point the initializer at that folder:

import Title

let titles = try await Titles(directory: modelFolder)
let card = try await titles.describe(text)

card.title          // "Filming a two-person podcast on iPhone"
card.description    // one or two sentences
card.isEmpty        // true when the model returned neither

On unusual input, Title can keep generating without end. maxTokens caps each run, at 96 tokens by default:

let titles = try await Titles(directory: modelFolder, maxTokens: 96)

With Clips

To give each clip a title and a description, pass the clips to Title. cards(for:) returns the cards in the same order as the clips.

let cards = try await titles.cards(for: moments)   // index-aligned with moments
let card = try await titles.card(for: moments[0])

cards(for:) writes one card at a time. One generation already keeps the GPU busy, so parallel generations aren't faster. On a phone, parallel generations also add heat, and the phone throttles partway through a long video.

Other kinds of text

We fine-tuned Title on transcript clips. Title also writes accurate cards for news paragraphs, product descriptions and emails. Title can still get details wrong on general prose.

Files

Title's files form one MLX model folder. Pass the whole folder to Titles(directory:).

File	Contents
`model.safetensors`	6-bit quantized weights
`model.safetensors.index.json`	Shard index. Keep this file: the loader needs it, even with a single shard.
`config.json`	Architecture and quantization config
`generation_config.json`	Decode defaults
`tokenizer.json`, `tokenizer_config.json`	Byte-level BPE with merges
`chat_template.jinja`	The chat template we trained the fine-tune against

Keep the chat template as shipped. A different template changes the task the model performs.

The prompt

The SDK sends Title the exact instruction we fine-tuned the model on, and parses the reply for you. The model replies with two labeled lines:

TITLE: <3-8 words, no final punctuation>
DESC: <1-2 sentences>

The model sometimes drifts off this format, so parse the reply tolerantly. When the reply has no labels, the SDK takes the first line as the title. When the SDK finds neither a title nor a description, card.isEmpty is true.

Apple only

MLX runs only on Apple silicon, so Title has no Android, Linux, Windows, browser or Node build. We also built Title for Core ML, to compare. The Neural Engine is limited by memory bandwidth when a model writes short text one token at a time. The Core ML build lost to MLX on time to first token, tokens per second, load time and memory use. We don't plan to ship the Core ML build.

Limitations

Title sometimes opens a description with a stock phrase that its instruction forbids. Read each card before it reaches a user.
Title can get details wrong on general prose. It's most accurate on transcript clips, news paragraphs, product descriptions and emails.
On unusual input, Title can keep generating without end. maxTokens caps each run.

Built on

ibm-granite/granite-4.0-350m: the base model we fine-tuned Title from.

License

Desert Ant Labs Source-Available License. Free for most apps, and a commercial license is required at scale. Full terms are at the link. Licensing: licensing@desertant.com.

Third-party notices for the weights: THIRD_PARTY_NOTICES.md.

Citation

@software{title_2026,
  title  = {Title: On-device titles and descriptions: a short factual title and a one- to two-sentence description for any passage of text},
  author = {Desert Ant Labs},
  year   = {2026},
  url    = {https://huggingface.co/desert-ant-labs/title},
}

Downloads last month: 276