Fetching from the wire…
Top 5 · 2026-03-07 · source-backed
AI2's fully open 7B model replaces 75% of attention layers with Gated DeltaNet linear recurrence, matching OLMo 3 accuracy on MMLU with 49% fewer tokens. Nathan Lambert's analysis highlights the theoretical contribution: hybrid models formally solve problems that neither pure transformers nor linear RNNs can solve alone. Combined with Phi-4-reasoning-vision-15B (selective reasoning at 15B params, trained on 200B tokens in 4 days), small open models are producing genuinely competitive alternatives to API-dependent workflows. Architecture choice matters again, not just scale. (AI2 Blog | Interconnects)
Each link below shares sources, entities, or timing with this story.
Shared entities / Same source / Shared topic
Both cover Combined, Interconnects, Nathan Lambert, OLMo Hybrid; cite the same source (Interconnects); overlapping topics (alone, architecture, combined, hybrid, model).
Shared entities / Same source / Shared topic / What happened next
Both cover AI2 Blog, Olmo, RNNs; cite the same source (AI2 Blog); overlapping topics (alone, architecture, hybrid, model, olmo).
Shared entities / Same source / What happened next / Tension
Both cover AI2, Olmo Hybrid; cite the same source (AI2 Blog); picks up the AI2 thread on 2026-03-08.
Shared entities / Same source / Shared topic
Both cover Interconnects, OLMo Hybrid; cite the same source (Interconnects); overlapping topics (hybrid, model).
Shared entities / Same source / What happened next
Both cover Interconnects, Olmo Hybrid; cite the same source (Interconnects); picks up the Interconnects thread on 2026-03-08.
Both cover AI2, Olmo Hybrid; cite the same source (AI2 Blog); picks up the AI2 thread on 2026-03-08.
Shared entities / Same source domain / Earlier coverage / Tension
Both cover Interconnects, Nathan Lambert; reported by the same outlet (interconnects.ai); earlier Interconnects coverage from 2026-03-06.
Sparse Delta Memory benchmarked against Gated DeltaNet / Shared entities / What happened next
Linked by a graph relationship (Sparse Delta Memory benchmarked against Gated DeltaNet); both cover Gated DeltaNet, RNNs; picks up the Gated DeltaNet thread on 2026-07-12.