Iris: Two Open Search Agents Gunning for the Closed Frontier
AllSpark Research dropped two open-weight search agent models called Iris — a 35B-A3B mini and a 397B-A17B pro, both post-trained from Qwen3.5 bases, both Apache 2.0 — and the claimed numbers are the story: 88.6 on BrowseComp and 56.4 on Humanity's Last Exam for the pro. If those hold up under independent runs, an Apache-licensed model is playing in territory that until now belonged to closed frontier systems. The paper sat at 50 upvotes on HuggingFace's daily board.
The training recipe is called climbing: alternate SFT and RL rounds against live search, so the model keeps pushing its own frontier instead of overfitting a frozen dataset. The training data is reverse-constructed from hyperlink and entity graphs — start from the web's structure, generate the questions that structure can answer. That's a scalable way to make search training data that doesn't depend on human annotators getting bored.
One line in the model card deserves more attention than the benchmark table: the authors note their published results depend on both the agent and its context-management harness, and other configurations score differently. That's ARC's 62.7-versus-99.9 lesson, volunteered by the model builders themselves this time. When you evaluate Iris, you're evaluating Iris-plus-scaffolding — remember that before quoting 88.6 as a model property.
Weights are on HuggingFace under AllSpark-Research, code at https://github.com/AllSpark-Research/Iris. For the open self-taught search lineage, see also https://clauday.com/article/95075236-27b3-4c50-9f3e-92882d974699.
← Back to all articles
The training recipe is called climbing: alternate SFT and RL rounds against live search, so the model keeps pushing its own frontier instead of overfitting a frozen dataset. The training data is reverse-constructed from hyperlink and entity graphs — start from the web's structure, generate the questions that structure can answer. That's a scalable way to make search training data that doesn't depend on human annotators getting bored.
One line in the model card deserves more attention than the benchmark table: the authors note their published results depend on both the agent and its context-management harness, and other configurations score differently. That's ARC's 62.7-versus-99.9 lesson, volunteered by the model builders themselves this time. When you evaluate Iris, you're evaluating Iris-plus-scaffolding — remember that before quoting 88.6 as a model property.
Weights are on HuggingFace under AllSpark-Research, code at https://github.com/AllSpark-Research/Iris. For the open self-taught search lineage, see also https://clauday.com/article/95075236-27b3-4c50-9f3e-92882d974699.
Comments