Agnes AI releases Agnes-3.0-Flash Preview, an open-weights model with 33B parameters and a 262,144-token context window under the Apache License 2.0
Agnes AI introduced Agnes 3.0 Flash Preview, an open weights preview checkpoint distinct from the production/API model, with 33B parameters and a 262,144 token…
Published on MyPrivateClaw
Sep 13, 2026, 1:16 AM UTC
Coverage date
Sep 12, 2026
Last updated
Sep 13, 2026, 1:16 AM UTC
News summary
Agnes AI introduced Agnes 3.0 Flash Preview, an open weights multimodal model with 33B parameters and a 262,144 token context window, released under the Apache License 2.0. The checkpoint is an open weights preview checkpoint, distinct from the production/API model, and requires trust remote code and transformers =5.12 for loading. Agnes 3.0 Flash Preview is specified as a hybrid attention decoder with 72 decoder layers, split as 54 gated delta rule recurrent layers and 18 global attention layers, so only 18 of the 72 layers hold a KV cache that grows with context. The Preview is listed with text, image and video understanding and tool calling. Reference benchmark results for Agnes 3.0 Flash Preview list 85.05 on GPQA Diamond, 74.20 on IFBench, 68.33 on AA LCR and 38.08 on SciCode. These are reference results compiled from different sources and snapshots, not a controlled head to head c…