SHELFCRITTER AI Digest

COMPANY

ai2

companytopic-noteopen-sourceresearch

Overview

ai2 (the Allen Institute for AI) is a Seattle-based non-profit research lab known for OLMo, Tülu, and other open-weights model releases. On May 9, 2026 ai2 released EMO, a 1B-active / 14B-total mixture-of-experts model trained on 1T tokens, with the structural distinction that experts are routed at the document level — clustering around domains (health, news, etc.) rather than surface patterns. The release is open-weights with the full collection on Hugging Face under allenai/emo.

Timeline

  • 2026-05-09-AI-Digest — ai2 releases EMO (1B-active / 14B-total MoE, trained on 1T tokens) on Hugging Face. The interesting structural choice is document-level routing: experts cluster around domains (health, news, etc.) rather than surface patterns. Most published MoE designs route per-token; document-level routing is closer to retrieval-augmented sparsity than to Mixtral-style per-token gating. Open-weights, full collection at allenai/emo.

Key Developments

  1. Document-Level Expert Routing: The substantive architectural choice — experts cluster around domains (health, news, etc.) at document granularity rather than per-token. Most published MoE designs route per-token; document-level routing is closer to retrieval-augmented sparsity than to Mixtral-style per-token gating. The structural angle is the news, not the absolute capability tier.

  2. EMO Release as Open-Weights MoE: 1B-active / 14B-total MoE, trained on 1T tokens, full collection at allenai/emo. Joins the broader 2026 open-weights MoE cohort (DeepSeek V4, Qwen, Llama 4 Scout) but at a smaller activated-parameter footprint than the frontier-tier Chinese-lab releases.

See also: EMO, MOC - Open Source Models.