33.09B (dense)
262,144 tokens
Apache License 2.0
Open-weight native multimodal reasoning model designed for coding, instruction following, long-context work, and tool-driven agents. Agnes-3.0-Flash accepts text, images, and video and generates text, with a 33.09-billion-parameter dense decoder and an approximately 66 GB BF16 checkpoint. Its hybrid architecture contains 72 decoder layers arranged in a repeating 3:1 pattern: 54 gated delta-rule recurrent layers and 18 global-attention layers, meaning only one quarter of the layers maintain a KV cache that grows with context. The model uses a 5,120-dimensional hidden state, 24 query heads and four KV heads in global-attention layers, causal-convolutional gated delta layers, SwiGLU feed-forward networks with an additional parallel 2,048-wide SwiGLU branch, and a 248,320-token vocabulary. Its vision tower has 27 layers, a 1,152-dimensional hidden state, 16-by-16 patches, 2-by-2 spatial merging, and projection into the language model's hidden space. The supplied chat template supports xhigh, medium, and low reasoning effort plus thinking-disabled operation, structured tool calling, and multimodal conversation. The checkpoint requires its bundled custom Transformers implementation with trust_remote_code enabled; Agnes also supplies patches and a launch script for SGLang. Agnes reports 74.20 on IFBench, 38.08 on SciCode, 85.05 on GPQA Diamond, 68.33 on AA-LCR, and 23.00 on AA Omniscience Accuracy, while cautioning that its comparison table combines different sources, harnesses, model snapshots, and parameter conventions rather than a controlled head-to-head evaluation. The official BF16 deployment guidance recommends one H100 80GB or H200 141GB and at least 128GB of host memory.