Ant Group's AI lab, Ant Ling (Inclusion AI), has officially unveiled its Ling-2.6-flash model and confirmed it's the same one that previously appeared anonymously on OpenRouter under the name Elephant Alpha. It's a 104-billion-parameter model with 7.4 billion activated parameters, built on a hybrid linear architecture for better computational efficiency. The whole thing is designed from the ground up for continuous agent workflows.
https://twitter.com/AntLingAGI/status/2046660999491858521
The goal was to fix a common problem with traditional models in multi-step agent tasks: they're computationally expensive and tend to produce overly verbose outputs. Ling-2.6-flash was specifically trained to generate concise responses. According to Ant Ling, in agent testing, it uses only about one-tenth the tokens of a model like Nemotron-3-Super, while its prefill throughput is 2.2 times higher. It can generate text at 340 tokens per second on a setup with four H20 GPUs and has achieved state-of-the-art results for its class on benchmarks like SWE-bench Verified and TAU2-bench.
The model is now available for a one-week free API trial on OpenRouter and Ant Ling's official platform. The lab says it plans to open-source the model along with BF16, FP8, and INT4 quantized versions. It's also partnering with Ant Digital Technology to launch a commercial version called LingDT. This move into the agent space comes as other players, like Moonshot AI with its open-sourced Kimi K2.6, are also pushing hard on coding and agent performance.