Ling-3.0-tiny

Ling-3.0-tiny

Generates text with a 7.9B hybrid reasoning model that activates only 1.3B parameters per token, so it fits local and low-resource setups. Suits developers who want reasoning and agents without datacenter GPUs.

Made by inclusionAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ling-3.0-tiny is a lightweight reasoning model from inclusionAI with 7.9B total parameters and 1.3B active for each token. It can answer fast or reason in several steps, and you switch thinking on or off per request. It ships in BF16, FP8 and INT4 weights for a range of hardware.

Who it suits: Developers who want reasoning, coding and agent abilities at a low inference cost, on local machines or modest hardware.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a small reasoning model locally for coding help | Serve it on a MacBook or a Mac mini
Agents | Agent builders | Use it as a cheap model inside agent workflows | Use thinking mode only for hard steps
Education | Students | Work through math and science problems on a laptop | Ask a step-by-step question with thinking on

If you need the largest model quality and have datacenter GPUs, you can skip it.

Where to use it:

• SGLang, which has a cookbook for this model
• NVIDIA DGX Spark
• Apple Silicon MacBooks and Mac minis
• Windows, macOS or Linux, for downloading the files

Use case | Who | What they get back
Local reasoning | Developers | Multi-step answers from a small model
Fast replies | App builders | Quick answers with thinking switched off
Agent tasks | Agent builders | Coding and tool-style task handling

How to set it up:

• Python with pip
• Hardware that fits the weights, about 15.8 GB in BF16 or about 3.9 GB at 4-bit by Hugging Face's estimate
• The Quickstart section of the model card, which covers SGLang

1. Install the Hugging Face hub package.
2. Download the model files.
3. Follow the card's Quickstart to serve it.

pip install -U huggingface_hub
hf download inclusionAI/Ling-3.0-tiny

How to use it: Send prompts through your serving stack. Thinking mode is on by default, and the enable_thinking setting turns it on or off for each request. For sampling, the card advises a temperature of 1.0, with top_p set to 0.95 and top_k set to 20.

Tool | What it does
enable_thinking | Turns thinking mode on or off for a request
SGLang | Serves the model, with a cookbook of launch recipes

Example: You send a short math question with thinking on. The model reasons through the steps first and then gives its answer. With thinking off, you get a quick reply.

For newcomers and experts: If you are new, run the FP8 or INT4 copy on a laptop and start with thinking off for quick questions. Switch it on for harder problems.

The architecture stacks 3 Kimi Delta Attention layers for every 1 Multi-Head Latent Attention layer, with a sparse mixture-of-experts feed-forward block of 128 routed experts. Each token uses 8 routed experts plus 1 shared expert. The card says this balances long-context ability, parameter efficiency and compute cost.

A token enters a 4-layer block
3 KDA layers and 1 MLA layer process it
The router picks 8 experts and 1 shared
Thinking mode is on or off per request

Good to know:

• The card reports FP8 speeds of about 100 to 105 tokens per second on DGX Spark
• It reports 86 to 90 tokens per second on an M4 Pro MacBook, with about 8.34 GiB of peak memory at 8K context
• License: MIT
• Cost to run: none named, you supply the hardware
• Data it reaches: your prompts and the files downloaded from Hugging Face
• Not for: users who need datacenter-scale quality from a small model

Title: Parameters in billions
Total: 7.9 billion
Active per token: 1.3 billion

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 7.9B
• Weights: about 15.8 GB in BF16; about 3.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 13,827
• Likes: 583
• License: mit
• Quantized versions: 41
• Fine-tunes: 16
• Tags: safetensors, bailing_hybrid, text-generation, conversational

Links

• Model: https://huggingface.co/inclusionAI/Ling-3.0-tiny
• Files: https://huggingface.co/inclusionAI/Ling-3.0-tiny/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/Ling-3.0-tiny

Then follow the usage in its model card. Nothing in this kit runs by itself.