Ant Group's Ling‑2.6‑1T Model

Ant Group has released Ling-2.6-1T, a trillion-parameter language model built to follow instructions quickly while using fewer tokens and less compute. (ant-ling.com) Large language models work by predicting the next chunk of text, and bigger models usually cost more to run because they generate longer answers and touch more of the system on each step. Ant’s Ling line uses a Mixture-of-Experts design, which keeps many parameters available but activates only a smaller slice for each token. (github.com) (huggingface.co) Ant has been pushing that design for months. Its open-source Ling-2.5-1T model, released in February 2026, had 1 trillion total parameters with 63 billion active parameters and a context window of up to 1 million tokens, according to its Hugging Face model card. (huggingface.co) The company’s own product pages frame the Ling series as general-purpose models for conversation, text generation, code generation, and content understanding, while the separate Ring series is tuned for deeper reasoning tasks. Ling-2.6-1T now sits in that general-purpose family alongside Ling-2.6-flash and Ling-2.6-mini. (developer.ant-ling.com) (ant-ling.com) That split tracks a broader argument inside artificial intelligence labs: some models spend more tokens “thinking” through a problem, while others aim to answer faster with less overhead. Ant’s earlier Ling-2.5-1T release said “instant models” expand reach by balancing efficiency and performance, while “thinking models” raise the ceiling of intelligence. (huggingface.co) Ant is not entering this race from the sidelines. The company said in October 2025 that it was open-sourcing Ling-1T, a trillion-parameter general-purpose model, and described its model family as three lines: Ling for non-thinking models, Ring for thinking models, and Ming for multimodal systems that handle text, image, audio, and video. (antgroup.com) The commercial angle is visible in Ant’s developer pricing. As of April 23, 2026, Ant listed Ling-2.6-flash at ¥0.60 per million input tokens and ¥1.80 per million output tokens, versus ¥4.00 and ¥8.00 for Ling-2.5-1T, with a temporary free-access offer for Ling-2.6-flash through April 28. (developer.ant-ling.com) Ant’s public materials also tie these models to real products. The company says its BaiLing foundation model already supports applications including Maxiaocai, a financial assistant for market analysis and portfolio diagnostics, and CodeFuse, an enterprise coding tool. (antgroup.com) The release adds another Chinese fintech player to a fast-moving contest over cheaper, faster instruction models for chatbots, coding tools, and software agents. Ant’s pitch with Ling-2.6-1T is straightforward: keep the trillion-parameter scale, but make the model answer more like a service product than a research demo. (ant-ling.com) (antgroup.com)