Climbing to the Search Frontier.
Iris-mini (35B-A3B) and Iris-pro (397B-A17B) are open-weight search agents post-trained from the Qwen3.5/3.6 series. A capable search agent has to decide what to search, how to read what comes back, when to keep going, and when the evidence it has gathered is enough. Iris is trained for exactly that loop.
 |
 |
 |
 |
30–35B
| Model | Size | BrowseComp | BrowseComp-ZH | DeepSearchQA | HLE |
|---|---|---|---|---|---|
| MiroThinker-1.7-mini | 30B | 67.9 | 72.3 | – | 36.4 |
| FORT-Searcher | 30B | 72.2 | 75.0 | – | – |
| Apodex-1.0-mini | 35B | 71.5 | 80.6 | 82.2 | 46.8 |
| Nex-N2-mini | 35B | 74.1 | 79.6r | 87.2r | 37.1r |
| Agents-A1 | 35B | 75.5 | – | – | 47.6 |
| XYZ-Aquila-mini | 35B | 78.8 | 82.9 | 89.5 | 51.1 |
| Iris-mini | 35B | 82.2 | 84.8 | 86.9 | 52.3 |
~400B
| Model | Size | BrowseComp | BrowseComp-ZH | DeepSearchQA | HLE |
|---|---|---|---|---|---|
| MiroThinker-1.7 | 397B | 74.0 | 75.3 | – | 42.9 |
| Apodex-1.0 | 397B | 75.5 | 82.6 | 84.6 | 49.0 |
| Nex-N2-Pro | 397B | 83.7 | 79.6r | 92.3r | 50.0r |
| XYZ-Aquila-pro | 397B | 84.8 | 85.1 | 92.5 | 53.3 |
| Iris-pro | 397B | 88.6 | 85.1 | 92.9 | 56.4 |
DeepSearchQA is scored with F1, the rest with accuracy; HLE uses the text-only subset. Iris numbers
use the discard-all context-management setting; baselines come from their public reports, each
under its own context management. r reproduced by the XYZ-Aquila team.
| Model | Base | Params (total / active) | Context | Download |
|---|---|---|---|---|
| Iris-mini | Qwen3.6-35B-A3B | 35B / 3B | 256K | 🤗 Iris-mini |
| Iris-pro | Qwen3.5-397B-A17B | 397B / 17B | 256K | 🤗 Iris-pro |
Long-horizon search runs out of context before a hard question is resolved, so every serious system carries some mechanism for this. It is worth enough that a single published number belongs to the agent and its harness together, which is why we report every benchmark in both regimes, under one tool set, one context limit, and one judge.
| Setting | BrowseComp | BrowseComp-ZH | DeepSearchQA | HLE |
|---|---|---|---|---|
| Iris-mini | ||||
| w/o | 64.7 | 72.3 | 81.0 | 43.2 |
| retry | – | 83.0 | 89.1 | 52.0 |
| discard-all | 82.2 | 84.8 | 86.9 | 52.3 |
| discard-all + retry | 85.9 | 85.1 | 89.9 | 52.4 |
| Iris-pro | ||||
| w/o | 72.6 | 76.8 | 86.4 | 50.8 |
| retry | – | 84.1 | 92.3 | 56.6 |
| discard-all | 88.6 | 85.1 | 92.9 | 56.4 |
| discard-all + retry | 90.3 | 85.1 | 93.4 | 56.6 |
discard-all clears the accumulated tool history and restarts from the question once the running
context crosses a threshold. retry restarts an episode that ended without a parseable answer, carrying forward a short
summary of what was already ruled out. We report discard-all as the headline setting even where adding
retry scores higher.
Iris-Harness/ is the harness behind every number above: the agent loop, the two
tools, the context-management strategies, the four benchmarks and the graders. It runs against any
OpenAI-compatible endpoint.
cd Iris-Harness && uv sync
uv run python data/prepare_data.py
bash scripts/run_eval.sh --base-url http://127.0.0.1:21234/v1 --llm-config iris-mini \
--benchmarks "browsecomp:0:1" --context-discard-threshold 131072Iris is built on open-source work, and we are grateful to the teams behind it:
If you find Iris useful in your research, please cite our paper:
@misc{liu2026irisclimbingsearchfrontier,
title={Iris: Climbing to the Search Frontier},
author={Ziyuan Liu and Hengqi Liu and Zichuan Wang and Yang Qin and Jiachen Liang and Xu Chu and Shaowei Chen and Yuantao Gu and Zhaokai Luo and Yao Hu and Mu Chuan},
year={2026},
eprint={2609.04304},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2609.04304},
}The data construction and training pipelines are coming soon.
Questions or collaboration: reach us at liuziyua22@mails.tsinghua.edu.cn.