Agnes AI Ships 33B Multimodal With 262K Context Under Apache 2.0
TL;DR
- Agnes AI released Agnes-3.0-Flash, a 33B open-weights multimodal model under Apache License 2.0 with a 262,144-token context window.
- The hybrid decoder uses 54 gated delta-rule recurrent layers and 18 global-attention layers, so only 18 of 72 layers hold KV cache.
- The bf16 checkpoint (~66 GB) runs on a single NVIDIA H200 141 GB or H100 80 GB and scores 85.05 on GPQA Diamond, 74.20 on IFBench.
Agnes AI has released Agnes-3.0-Flash, a 33-billion-parameter open-weights multimodal model under Apache License 2.0, with a 262,144-token context window and support for text, image and video.
The architecture is the pitch. Of 72 decoder layers, 54 use a gated delta-rule recurrent design and 18 are standard global attention, so the model card notes that "Only 18 of 72 layers hold KV cache" as context grows. Agnes AI frames the trade as "flagship-class reasoning without flagship-class hardware," saying the bf16 checkpoint runs on a single NVIDIA H200 141 GB or H100 80 GB. Weights on disk are roughly 66 GB.
On benchmarks the card is candid. Agnes-3.0-Flash posts 85.05 on GPQA Diamond, 74.20 on IFBench, 68.33 on AA-LCR and 38.08 on SciCode. Those numbers sit consistently below the frontier comparisons the card places next to them, including 90.8 on GPQA Diamond and 50.3 on SciCode for DeepSeek V4, and 51.4 on AA-Omniscience for Gemini 3.5 Flash against Agnes's 23.00. The card describes the results as "Competitive across core capabilities."
It joins a busy run of open-weight drops we've been logging on our open-source tracker.
Originally reported by huggingface.co
Read the original article →Original headline: Agnes-AI Ships Open-Weight 33B Multimodal Agnes-3.0-Flash With 262K Context Under Apache 2.0