Ant Group has officially unveiled Ling 2.6 Flash, a new large language model that has scored 26 on the Artificial Analysis Intelligence Index, marking a 10-point improvement over its predecessor, Ling-flash-2.0. This model, developed by Ant Group's inclusionAI lab, emphasizes non-reasoning capabilities and a favorable cost-to-intelligence ratio, positioning it as a highly efficient option for AI agent applications. The announcement highlights a shift in the AI model competition towards "intelligence efficiency" rather than solely parameter scale.
The Ling 2.6 Flash model, utilizing a sparse Mixture-of-Experts (MoE) architecture with 104 billion total parameters and 7.4 billion active parameters, demonstrates notable token efficiency. According to Artificial Analysis, it consumed approximately 15 million output tokens to run the full Intelligence Index, a fraction of the 78 million tokens used by Qwen3.5 9B or over 110 million by Nemotron-3-Super. This efficiency translates to an estimated 86% reduction in inference cost for developers and enterprises, with the full index evaluation costing only around $23.
Performance gains in Ling 2.6 Flash were primarily driven by substantial improvements in agentic capabilities and instruction following. The model saw its τ²-Bench score jump from 21% to 86% (+65 points) and IFBench from 34% to 57% (+23 points), alongside an 84% increase in GDPval-AA Elo. Despite these advancements, its AA-Omniscience performance, which measures knowledge recall, registered a 96% hallucination rate, consistent with models having smaller active parameter counts.
Ant Group plans to make Ling 2.6 Flash open weights shortly after its release, though the weights are not yet available on Hugging Face. The model is accessible via a third-party API through Novita API, with pricing set at $0.10 per million input tokens and $0.30 per million output tokens. Previously known as "Elephant Alpha" during its testing phase on OpenRouter, Ling 2.6 Flash has a context window of 262K tokens, double that of Ling-flash-2.0, and is part of Ant Group's broader AI family which includes the reasoning-focused Ring series and multimodal Ming series.
