I benchmarked 4 major LLM techniques from late 2025 / early 2026 for a local model on RTX 5070 (12GB VRAM). Here are the key takeaways: ## Mamba-3 (GoombaLab, March 2026) Tri Dao’s latest SSM overhaul in 3 ways: more general recurrence from classical control theory, data-dependent RoPE embeddings, and improved discretization. Two variants — SISO (Mamba-2 shapes) and MIMO (+1% accuracy at 1B).