Ant Group's small model claims to outperform its 1-trillion-parameter big brother

Ant Group's 124B Ling-3.0-flash model outperforms its 1T flagship on key benchmarks using a hybrid MoE architecture, now free on OpenRouter.

Five billion active parameters per token out of one hundred twenty-four billion in total: this is the ratio at which InclusionAI, the AI lab affiliated with Ant Group, positions Ling-3.0-flash, a hybrid reasoning MoE designed for production agents. The lab claims results equivalent to or better than those of its own one-trillion-parameter flagship model on most of the presented benchmarks, with one-eighth of the total parameters and one-twelfth of the active parameters.

The architecture is based on a native hybrid linear attention: KDA and MLA layers stacked in a five-to-one ratio. The former handle long-range memory at a lower cost, while the latter preserves the recall accuracy that purely linear models lose. Activating one expert out of sixty-four further lightens the computation. The native context reaches up to 256K tokens, with an announced margin toward one million, and a switchable reasoning mode allows toggling between thinking and direct response depending on the difficulty of the task.

The online demonstrations cover building a 3D city via Blender MCP, from the Python script to the aerial video rendering; an autonomous research team where five agents formulate hypotheses, cross-reference sources, and produce a paper accompanied by slides; and an office workflow utilizing Office MCP to format a Word proposal and verify Excel spreadsheets.

The model is accessible on OpenRouter, free of charge until August 3, 2026. The weights are not published at this stage, unlike previous versions of the series, with InclusionAI mentioning a later release.