Title

Names what you just wrote.

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

Install

Swift (requirements)

.package(url: "https://github.com/Desert-Ant-Labs/desert-ant-core.git", from: "3.1.0",
        traits: ["MLX"])

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

Usage

Apple only, and the one model here that runs on MLX rather than Core ML: short autoregressive decode measured 5.7-8.3x faster on the GPU than on the Neural Engine. That is why it is behind the MLX package trait, and why a build that forgets the trait fails at compile time instead of mis-building.

Loading is expensive and generation is cheap, so build one Titles and reuse it. It is an actor because MLX state is not safe to drive from several tasks at once.

Swift

Nothing is downloaded here. Point the initializer at a folder you populated with the model files, unlike the other models in this repo:

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

maxTokens caps a degenerate run, which is a real failure mode for a small instruct model given unusual input:

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

With Clips

Card is owned by this module and is not a field on Clip: selection and card writing are separate stages on separate silicon. Pair them when you want both.

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

cards(for:) is sequential on purpose. Decode is already GPU-bound, so overlapping generations contend for the same device rather than adding throughput, and on a phone it adds thermal pressure that shows up as throttling partway through a long video.

Not clip-specific

The model was fine-tuned on transcript clips, but the task it learned is general: news paragraphs, product descriptions and emails all produce accurate, correctly-registered cards. Treat it as capable on general prose, not infallible on it.

Files

An MLX model directory. Load the folder, not a single file.

The chat template is not incidental. A different template is a different task to this model.

The prompt

The model was fine-tuned against one specific instruction, and a paraphrase is a different task to it. It lives in Titles.prompt in the SDK; use that wording. The reply is two labelled lines:

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

Parse tolerantly. A card model that drifts off format should degrade to a usable title rather than throw.

Apple only

MLX runs on Apple silicon and nowhere else, so there is no Android, Linux or Windows artifact here and no manifest promising one. A Core ML export exists in the training repository and is kept as evidence rather than as a candidate: on short autoregressive decode the Neural Engine is bandwidth-bound, and the Core ML arm lost on first token, throughput, load time and resident memory.

Status

Internal testing, and less settled than that phrase usually implies. This card carries no quality figures: no independent review has been completed, and a known open issue is that the model sometimes opens a description with a stock phrase its own instruction forbids. Treat the output as needing a read before it reaches a user.

Built on

• ibm-granite/granite-4.0-350m: the base model this is fine-tuned from.

See THIRD_PARTY_NOTICES.md.