Ling-1T
Trillion-parameter efficient reasoning model by InclusionAI
Ling-1T is a one-trillion parameter language model from InclusionAI, released as part of the Ling 2.0 series. It uses a mixture of experts architecture to activate only fifty billion parameters per token, aiming for high efficiency in complex reasoning tasks.
- Contains 1 trillion total parameters with 50B active per token
- Trained on over 20 trillion high-quality reasoning-dense tokens
- Supports context lengths up to 128K tokens
- Uses FP8 mixed-precision training for efficiency
- Achieves ~70% tool-call accuracy on BFCL V3 benchmark
What it is
Ling-1T is a large language model developed by InclusionAI as the flagship entry in its Ling 2.0 series. It contains one trillion total parameters but activates only about fifty billion for each token processed. This design targets efficient reasoning and scalable cognition without requiring excessive computational resources during inference. The model supports a context window of up to one hundred and twenty-eight thousand tokens and handles text, visual understanding, and tool use.
How it came about
The model was built on the Ling 2.0 architecture, which was designed specifically for trillion-scale efficiency. InclusionAI guided its development using the Ling Scaling Law to ensure stability at massive sizes. The training process involved over twenty trillion high-quality tokens, with a significant portion dedicated to reasoning-dense data in later stages. Mid-training incorporated curated chain-of-thought corpora to enhance reasoning stability before final alignment.
Inside the model
The architecture uses a mixture of experts approach with a one in thirty-two activation ratio. It employs sigmoid-scoring expert routing and QK normalization to maintain stable convergence. Training utilized FP8 mixed precision, which the developers claim offers a fifteen percent speedup over BF16 while keeping loss deviation minimal. A custom learning rate scheduler called WSM helped simulate decay and boost generalization. Post-training used an evolutionary chain-of-thought process and a sentence-level policy optimization method known as LPO.
What it can do
InclusionAI states that Ling-1T performs strongly in code generation, mathematics, and logical reasoning. It reportedly ranks first among open-source models on the ArtifactsBench for visual aesthetics and front-end code generation. The model also demonstrates emergent tool-use capabilities, achieving around seventy percent accuracy on the BFCL V3 benchmark with minimal instruction tuning. It can interpret natural language instructions to create functional visual components and marketing copy.
How to use it
The model weights are available for download via Hugging Face and ModelScope, particularly for users in mainland China. Developers can run inference using frameworks like SGLang or vLLM, supporting both BF16 and FP8 precisions. An API is also provided through the ZenMux platform, allowing users to interact with the model via standard chat completion endpoints without managing local infrastructure.
Questions people ask about Ling-1T
How many parameters does Ling-1T have?
Ling-1T has one trillion total parameters. However, it uses a mixture of experts architecture that activates only about fifty billion parameters for each token processed, making it more efficient than dense models of similar size.
What is the context length of Ling-1T?
The model supports a context length of up to one hundred and twenty-eight thousand tokens. This allows it to process and reason over very long documents or extended conversations without losing track of earlier information.
Can I download Ling-1T for local use?
Yes, the model weights are available for download on Hugging Face and ModelScope. You can deploy it locally using inference engines like SGLang or vLLM, which support both BF16 and FP8 precision formats.
Good reading and sources
There is no English Wikipedia article on Ling-1T yet, so the facts on this page come from the sources listed above: huggingface.co/inclusionAI/Ling-1T. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.