Editorial illustration for Iris-mini and Iris-pro Lead Open-Weight Search Agent Class
Iris-mini and Iris-pro Open-Weight Search Agents
AllSpark, a Chinese AI lab, has released two open-source search agents, Iris-mini and Iris-pro, along with the full recipe used to train them. Iris-mini runs on 35 billion parameters, Iris-pro on 397 billion, and both are built on Qwen-series models with a 256,000-token context window. The team says both versions post the strongest results among open-weight search agents in their respective size classes.
The release matters because of an open question in this corner of AI research: when a search agent answers a question, how much is it actually researching versus just confirming facts it already absorbed during training. On the benchmarks most labs use to measure these systems, that distinction gets blurry. AllSpark's paper claims something else stands out too. The training data built for Iris-mini and Iris-pro seems to carry over to tasks the models were never explicitly trained on, including general tool use and office-style work.
That claim rests on how the training questions themselves get built, which is where AllSpark's method departs from the usual approach.
In a new paper, Chinese lab AllSpark describes two search agents of different sizes. Iris-mini has 35 billion parameters, and Iris-pro has 397 billion. Both build on Qwen-series models (Qwen3.6-35B-A3B and Qwen3.5-397B-A17B), work with a 256,000-token context window, and deliver the strongest results among open-weight search agents in their respective size class, according to the team.
Why this matters
The open-weight framing is the part worth watching. AllSpark didn't just publish two models, they published the training recipe, which means anyone can inspect how the reward model was built and whether that judge-derived-from-data approach actually holds up outside their benchmarks. That's a real test for the "trained on search, generalizes to office work" claim in the paper.
We've seen plenty of papers assert transfer learning gains without much scrutiny of the eval design, so the fact that AllSpark is giving researchers the actual recipe rather than just weights and a leaderboard number is what makes this worth a second look, not the "strongest in its class" tagline. For developers building agents that need to reason about when to search, when to stop, and when a loop is going nowhere, the two-part verification approach (path-checking plus judge review) is a more interesting engineering detail than the benchmark scores themselves. Whether it holds up under independent replication is the actual question here, not the headline.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives UC Berkeley Law Bans AI From Graded Work in Skills Debate
From Archives Altman, Musk Back Amodei's AI Oversight Call, Says OpenAI Won't IPO
Common Questions Answered
What are the key differences between Iris-mini and Iris-pro in terms of model size and architecture?
Iris-mini operates with 35 billion parameters based on Qwen3.6-35B-A3B, while Iris-pro uses 397 billion parameters based on Qwen3.5-397B-A17B. Both models are built on Qwen-series foundations and share a 256,000-token context window, allowing them to process significantly longer documents than many competing search agents.
Why is AllSpark's release of the training recipe significant for open-weight search agents?
By publishing the complete training recipe alongside the models, AllSpark enables researchers to inspect how the reward model was constructed and verify whether the judge-derived-from-data approach generalizes beyond their internal benchmarks. This transparency addresses a critical gap in AI research where transfer learning gains are often asserted without thorough scrutiny of evaluation design.
How do Iris-mini and Iris-pro perform compared to other open-weight search agents?
According to AllSpark's claims, both Iris-mini and Iris-pro deliver the strongest results among open-weight search agents in their respective size classes. This positions them as leading models in the open-source search agent category, though independent verification of these benchmarks remains important.
What is the connection between search training and the 'office work' generalization claim mentioned in the paper?
AllSpark's paper asserts that models trained specifically on search tasks can transfer their capabilities to broader office work applications. However, the article notes this claim requires scrutiny, as many papers make similar transfer learning assertions without rigorous evaluation design, which is why the public release of training methodology is particularly valuable for verification.
Further Reading
- Iris: Climbing to the Search Frontier - arXiv
- AllSpark-Research/Iris: Climbing to the Search Frontier. - GitHub
- Iris-mini und Iris-pro: Stärkste offene Suchagenten ihrer Klasse stehen jetzt als Open-Weight-Modelle bereit - The Decoder
- AllSpark 开源搜索 Agent Iris：BrowseComp 88.6 分刷新开源纪录——SFT-RL「攀登」配方全解 - UU AI Hub
- オープンソース検索エージェント「Iris」、複数ベンチマークで最高水準の性能を論文報告 - AI News JAPAN
