Description
Ling is a family of mixture-of-experts (MoE) large language models developed and open-sourced by inclusionAI, a series of open-source AI projects from Ant Group. The GitHub repository hosts the code, documentation, and links to model weights, and is actively maintained, with commits as recent as this month and 267 stars.
Read More
The project ships models in two main sizes: Ling-lite, with 16.8 billion total parameters and 2.75 billion activated parameters per forward pass, and Ling-plus, with 290 billion total parameters and 28.8 billion activated. Because Ling uses a mixture-of-experts architecture, only a subset of parameters activates for any given input, which keeps inference cost closer to a much smaller dense model while retaining a larger overall parameter count. Ling-lite supports a 128K token context window; Ling-plus supports 64K. A coding-focused variant, Ling-Coder-lite, is also available with a 16K context window.
Model weights are distributed through Hugging Face and ModelScope (the latter aimed at users in mainland China for faster downloads), with both base and instruction-tuned checkpoints for most variants. The repository documents benchmark results against comparable open models such as Qwen3 and LLaMA 3.1 across tasks including MMLU, HumanEval, LiveCodeBench, and math reasoning benchmarks, where Ling-lite-1.5 shows competitive or leading scores on several coding and math benchmarks among similarly sized models.
For deployment, Ling includes documented paths for running inference through the Hugging Face Transformers library, vLLM, and Huawei's MindIE, plus example code for both offline batch inference and online serving, including multi-machine setups for the larger Ling-plus model. The repository also includes fine-tuning instructions for teams that want to adapt the models to their own data.
Ling is released under the MIT license, and the code is free to use, modify, and redistribute. It's aimed at researchers, ML engineers, and developers who want to self-host or customize an open LLM rather than call a hosted API, particularly those interested in mixture-of-experts architectures or coding-specific model variants.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsWeb
Ling's Core Features
Mixture-of-experts architecture for efficient inference at scale
Multiple model sizes: Ling-lite (16.8B) and Ling-plus (290B) total parameters
Ling-Coder-lite variant tuned for coding tasks
Context windows up to 128K tokens (Ling-lite) or 64K (Ling-plus)
Base and instruction-tuned checkpoints for most variants
Weights distributed via Hugging Face and ModelScope
Deployment support for Transformers, vLLM, and MindIE
Documented fine-tuning workflow for custom training
MIT-licensed, open-source code
Getting Started with Ling
Clone the repo: get the code from github.com/inclusionAI/Ling
Download weights: pull the model checkpoint you need from Hugging Face or ModelScope
Install dependencies: set up Transformers, vLLM, or MindIE per the repo's deployment guides
Run inference: use the provided quickstart code for offline or online inference
Fine-tune (optional): follow the finetuning guide to adapt the model to your own data
Evaluate: compare outputs against the documented benchmark results for your use case
Ling's Use Cases
- Self-hosted LLM inference
- Coding assistance research
- Custom fine-tuning
- Long-context tasks
