- AIAnt Group has open-sourced Ling-2.6-flash, offering multiple quantized versions to meet diverse hardware environments and enterprise deployment needs.
- The model features high inference efficiency and low token consumption, aiming to significantly reduce the costs of enterprise-grade agent applications.
Ant Group's Ling large model has open-sourced Ling-2.6-flash, marking another significant move by the company in the field of enterprise-grade artificial intelligence (AI) applications.
The instruct model, which features 104 billion total parameters and 7.4 billion active parameters, was officially made available for global developers to download on Wednesday, according to an announcement on Wednesday.
Two weeks ago, the model made its debut on the OpenRouter testing platform under the anonymous identity "Elephant Alpha," quickly drawing widespread attention from the developer community.
Now officially open-sourced, multiple versions including BF16, FP8, and INT4 have been simultaneously provided for flexible deployment.
The core technical advantage of Ling-2.6-flash lies in its innovative hybrid linear architecture, which allows the model to fundamentally optimize overall computational efficiency from the ground up, according to the announcement.
Under a 4x H20 hardware setup, its inference speed can reach up to 340 tokens per second, significantly reducing costs.
During the training process, the model was specifically calibrated for token usage efficiency, striving to accomplish complex predefined tasks with more concise text outputs.
In the full evaluation by Artificial Analysis, its token consumption was only about one-tenth of models like Nemotron-3-Super, substantially boosting its overall intelligence-to-efficiency ratio.
Targeting the currently high-demand agent application scenarios, the research and development team has continuously improved its capabilities in tool calling, multi-step planning, and complex task execution.
Based on real feedback collected over the past two weeks, the team has further enhanced its natural bilingual switching between Chinese and English, and improved its compatibility with coding frameworks.
Through this comprehensive open-source initiative, Ant hopes to explore a new paradigm for highly efficient, low-cost, and highly practical large model applications alongside global developers, the statement said.
This move will not only accelerate the commercial popularization of agent workflows but also provide a cost-effective technical option for enterprise-grade AI applications.
