Agnes-3.0-Flash Actually Refers to Two Models: Open-Weight and Leaderboard Results Must Not Be Conflated

Agnes-3.0-Flash quickly attracted attention in the local-model community after its release, but the latest model card clarifies a crucial versioning issue: Hugging Face provides an earlier 33B Preview checkpoint, while the model listed by Artificial Analysis is a production checkpoint served through the Sapiens AI API, with undisclosed weights and parameter count. They share the same name but are not the same model.

The open version uses a 72-layer hybrid decoder. Of those layers, 54 use gated delta rule recurrent state, while only 18 use global attention, meaning that the KV cache grows with sequence length in just one-quarter of the layers. It also features a hidden size of 5,120, a 262,144-token context window, a vision tower, and support for text, image, and video input. Its BF16 weights occupy approximately 66GB, and the model card recommends at least an 80GB H100 or a 141GB H200. Loading it also requires `trust_remote_code=True`, which means the Python implementation included with the repository should be reviewed before deployment.

The API version, meanwhile, claims a 1-million-token context window. Artificial Analysis labels it proprietary and provisionally assigns it an Intelligence Index score of 36. Provider-side testing measured approximately 234.8 output tokens per second and a time to first token of 1.82 seconds; the page also states that a full independent evaluation is still pending. These figures cannot be applied to the publicly available weights. The Preview checkpoint’s own reported results—including 85.05 on GPQA Diamond and 38.08 on SciCode—were compiled from different harnesses, snapshots, and sources, and the model card warns that they do not constitute controlled, head-to-head comparisons.

What engineering teams should focus on is not the naming dispute itself, but the need to include precise model identification in evaluation and deployment records. At a minimum, teams should retain the checkpoint revision, weight hash, remote-code revision, context configuration, and actual API provider. The next question is whether Agnes AI will disclose the production checkpoint’s architecture, parameter count, and reproducible evaluations, or allow third parties to test the current Preview weights directly. Until then, “the open 33B model scored 36 on Artificial Analysis” is not a valid technical conclusion.