Model reference · open weights

title

Not for GPU servers LLMs desert-ant-labs Text gen 1 build Its own licence terms 974 dl/mo

title is an open-weight language model from desert-ant-labs. Built for Apple silicon (MLX). It does not run on NVIDIA GPUs.

  • Title is a 352M parameter multilingual text-generation model by Desert Ant Labs that generates short factual titles and one- to two-sentence descriptions for any passage of text.
  • It supports a 32,768 token context length and is designed for on-device use on Apple platforms using MLX.
  • The model is built on ibm-granite/granite-4.0-350m and is distributed under the Desert Ant Labs Source-Available License.

Summary of the desert-ant-labs/title model card, 2026-10-01

What it is

Released bydesert-ant-labs
Released2026-08-09
Parameters352M
VRAM286 MB for the weights

What it runs on

Memory and cards for title (BF16)

286 MBweights, file size
29 MBcache per 1K tokens
566 MBruntime overhead, at least
32,768 tokenscontext max
CardRequests at onceContext maxMemory
8K each32K each
RTX 3060 12 GB4511all 32K11.6 GB
RTX 4060 Ti 16 GB6115all 32K15.4 GB
RTX 3090 24 GB9624all 32K23.4 GB
RTX 4090 24 GB9523all 32K23.4 GB
RTX 5090 32 GB12832all 32K31.0 GB
L40S 48 GB18345all 32K44.0 GB
A100 80 GB32982all 32K78.2 GB
H100 80 GB31478all 32K78.1 GB
RTX PRO 6000 Blackwell 96 GB38195all 32K93.8 GB
DGX Spark (GB10) 128 GB unified439109all 32K107 GB
H200 141 GB569142all 32K138 GB
B200 180 GB731182all 32K176 GB
Memory needed at each load
Requests at once8K tokens each32K tokens each
11.1 GB1.8 GB
52.0 GB5.5 GB
82.7 GB8.4 GB
164.6 GB15.9 GB
328.4 GB30.9 GB
6415.9 GB61.0 GB

One card, with vLLM's small-card settings.

From the model card

What desert-ant-labs says about title

Read the model card

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.

  • SDKs, install and examples: https://github.com/Desert-Ant-Labs/desert-ant-core/blob/main/docs/models/title.md
PlatformsiOS, macOS, tvOS, visionOS
Weightsv0.1.0

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.

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

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 is meant to identify this passage rather than its topic.

Files

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

FileContents
model.safetensors6-bit quantized weights
model.safetensors.index.jsonshard index; present even for one shard, because the loader reads it
config.jsonarchitecture and quantization config
generation_config.jsondecode defaults
tokenizer.json, tokenizer_config.jsonbyte-level BPE with merges
chat_template.jinjathe chat template the fine-tune was trained against

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:
DESC:

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

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: .

See 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},
}

© 2026 Desert Ant Labs ·

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

Running it

Where it runs

Built for Apple silicon (MLX). It does not run on NVIDIA GPUs.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms