Iris README

The project's README from github.com/AllSpark-Research/Iris. As one markdown file: README.md.

🤗 Hugging Face  |   💻 GitHub  |   📄 Tech Report

Iris

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.

Performance

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.

Models

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

Context Management

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.

Evaluation

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 131072

Acknowledgements

Iris is built on open-source work, and we are grateful to the teams behind it:

Citation

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 [email protected].

Iris (AllSpark-Research/Iris)

Searching Database…

102 ·GitHub
Iris
What is this?Iris is a pair of open-weight search agent models (Iris-mini, 35B-A3B, and Iris-pro, 397B-A17B) post-trained from the Qwen3.5/3.6 series to decide what to search, how to read results, when to keep searching, and when enough evidence has been gathered.
Is this a trained model or a framework?Both: Iris is a pair of released model checkpoints, and the repo also ships Iris-Harness, the agent loop, tools, and context-management code used to run and evaluate them.
How do I get it running?Download the Iris-mini or Iris-pro weights from Hugging Face, then run them through Iris-Harness, which works against any OpenAI-compatible endpoint.
Do I need a specific base model to use Iris?Iris is post-trained on top of Qwen3.6-35B-A3B (for Iris-mini) and Qwen3.5-397B-A17B (for Iris-pro), so it inherits those base architectures.
Can I reproduce how Iris was trained?Not yet - only the model weights and evaluation harness are out; the data construction and training pipelines are still unreleased.
What is the 'discard-all' context management setting?It clears the accumulated tool-call history and restarts the agent from the original question once the running context crosses a set threshold, used as Iris's headline reported setting.
What benchmarks is Iris evaluated on?BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE (Humanity's Last Exam, text-only subset), with DeepSearchQA scored by F1 and the rest by accuracy.
What's the difference between Iris-mini and Iris-pro?Iris-mini is a 35B parameter model with 3B active (built on Qwen3.6), while Iris-pro is a much larger 397B model with 17B active (built on Qwen3.5); both support a 256K context window and Iris-pro scores higher across every benchmark.
What open-source projects does Iris build on?The authors credit MiroThinker, Relax, ms-swift, and slime as the open-source foundations Iris was built on.
What are similar projects?RSIGame, RSI-Anything, Kimi-K3, MSA, NeoHorse, jeeves, DeepAnalyze, ZDTaichu5.0-9B
Which topics is it in?context-window, deep-research, mixture-of-experts, reasoning, llm-agents, autonomous-agents, search, qwen, evaluation, academic-research, large-language-models, agentic-ai
Creator
AllSpark-Research
Size
1.4 MB
Stars
102
Updated
CommitSep 28
Age
1mo
Commits · 90 Days