inclusionAI Releases Ling 2.6 1T, a Trillion Parameter Open Source Language Model

inclusionAI has released Ling 2.6 1T, a trillion parameter open source language model optimized for complex real world tasks such as coding and multi step agent workflows. The model was announced on Hugging Face and introduces several improvements in inference efficiency, token usage, and reasoning performance.

Ling 2.6 1T uses a hybrid architecture that combines Multi Head Latent Attention and Linear Attention to reduce latency and memory use for long contexts. It supports a context length of up to 262,144 tokens and achieves lower per token computational costs while maintaining high throughput.

The model introduces a Contextual Process Redundancy Suppression reward strategy that reduces unnecessary chain of thought reasoning, allowing faster and more concise responses. It integrates with frameworks such as Claude Code, OpenClaw, OpenCode, and CodeBuddy for coding and agent workflows.

On benchmark tests, Ling 2.6 1T achieved an Intelligence Index of 34 with about 16 million output tokens and led open source models on reasoning and execution tasks including AIME26 and SWE bench Verified. It also showed strong performance on long context and constraint handling benchmarks such as MRCR and IFBench.

Ling 2.6 1T is released under the MIT License and is available through OpenRouter and ZenMux for API access. Future updates aim to improve token efficiency, long range consistency, and multilingual alignment.