Motif 3: Technical Report

Motif Technologies

Research official 2 src. ~1 min

Technical report for Motif 3, a 314B-parameter (13.2B active) decoder-only Mixture-of-Experts model built on Grouped Differential Latent Attention plus manifold-constrained hyper-connections and a multi-token-prediction head, with competitive results on reasoning, coding, and long-context tasks.

Why it matters

20 upvotes on HuggingFace Daily Papers; a fully in-house architecture (not a Llama/Qwen derivative) from a smaller lab reaching frontier-adjacent agentic benchmark scores.

Importance: 2/5

Notable paper (20 HF Daily Papers upvotes, below the 100-upvote bump threshold).

mixture-of-experts architecture long-context

Sources

official Motif 3: Technical Report

secondary HuggingFace Daily Papers, 2026-08-11