MambaVision: A Hybrid Mamba-Transformer Vision Backbone

Pure Mamba backbones for vision underperformed ViT at high resolution, because vision needs precise local spatial detail that a global fixed-state recurrence discards.

Redesigns the Mamba formulation for visual features and, through ablation, establishes the actual layer structure: put Mamba-style selective-SSM blocks in the early stages and add self-attention blocks in the FINAL stages of each stage. The finding is that equipping Mamba with self-attention blocks in the last layers greatly improves its ability to capture long-range spatial dependencies. The resulting family is a hierarchical pyramid (patch embedding, then four stages of 2x downsampling) mixing Mamba blocks with a final self-attention block per stage. For serving this matters in the same way as every other hybrid: the attention layers reintroduce a KV cache for vision tokens, and because visual token counts are large, that cache is big. The state in the Mamba layers remains constant-size.

CVPR 2025 per the arXiv comment field; no journal-ref on the abs page. Recorded primarily as a vision-architecture datapoint in an inference-focused knowledge base: the Mamba-first/self-attention-last rule it establishes is now the standard hybrid recipe and reappears in the text hybrids (Jamba, Granite 4.0, Qwen3-Next). The paper reports ImageNet classification and downstream detection/segmentation accuracy and throughput at matched parameter count rather than a speedup ratio, so speedup_reported is null rather than reconstructed. MambaVision ships in NVIDIA's own inference stacks (NeMo / TAO / Triton model zoo) rather than as a general LLM-engine architecture, which is why adoption is in-upstream-engine rather than in-production broadly. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: cvpr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.