SHELFCRITTER AI Digest

MODEL

Iris

modeltopic-noteopen-sourcesearch-agent

Overview

Iris is a pair of open-weights search agents — Iris-mini (35B-A3B) and Iris-pro (397B-A17B) — trained by alternating SFT and RL against live search, with training tasks reverse-constructed from web hyperlink graphs. Introduced in the arXiv preprint Iris: Climbing to the Search Frontier (arXiv:2609.04304), the release reaches the strongest open-source results in its parameter classes on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. Iris is the latest and largest entry in an existing open lineage that includes Search-R1, Search-o1, R1-Searcher, DeepResearcher, ZeroSearch, and WebAgent-R1 — the load-bearing novelties are the frontier scale (397B-A17B) and the reverse-hyperlink task-synthesis pipeline, not a rare open recipe for agentic search.

Timeline

  • 2026-09-07-AI-Digest — Iris paper landed on the HuggingFace papers pass (▲12) as two open-weights search agents — Iris-mini 35B-A3B and Iris-pro 397B-A17B — trained by alternating SFT and RL against live search, with training tasks reverse-constructed from web hyperlink graphs. Reaches the strongest open-source results in its parameter classes on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. Load-bearing softener: Iris is not the rare open recipe for agentic search — Search-R1, Search-o1, R1-Searcher, DeepResearcher, ZeroSearch and WebAgent-R1 all predate it; what’s genuinely new is the frontier scale (397B-A17B) and the reverse-hyperlink task-synthesis pipeline. Carry as latest and largest in an existing open lineage, not as rare open recipe.

Key Developments

  1. Frontier-Scale Open Search Agent With Reverse-Hyperlink Task Synthesis (September 7, 2026): Iris-pro at 397B-A17B is the frontier-scale entry in an already-existing open lineage of RL-trained search agents. The load-bearing methodological novelty is the reverse-hyperlink task-synthesis pipeline — training tasks are reverse-constructed from web hyperlink graphs, letting the SFT+RL alternation train against live search without hand-authored task inventories. Strongest open-source results in class on BrowseComp / BrowseComp-ZH / DeepSearchQA / HLE. 30 / 60 / 90-day watch: independent replication of the reverse-hyperlink pipeline on a non-Iris backbone; whether Iris-pro weights land on HuggingFace at frontier scale under a permissive licence; whether closed-source search agents cite the pipeline in follow-on work.

See also: MOC - Open Source Models.