On September 7, 2026, the Iris team released two search proxy models: Iris-mini and Iris-pro. Iris-mini was trained on a dataset of 35B-A3B size, while Iris-pro was trained on a dataset of 397B-A17B size. The study proposed a training process called “SFT-RL climbing”: first, tasks were constructed back from the web hyperlink structure and data was filtered; after supervised fine-tuning (SFT), reinforcement learning (RL) optimization was performed using a reward discriminator and an observation summarizer within the training cluster. Ultra-long inferences were interrupted at the request layer to retain promised prefixes. Evaluation included benchmarks such as BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE. All results came from a single ReAct proxy, and context management functions were enabled when inference was performed. With this configuration, Iris-pro achieved scores of 88.6/85.1/92.9/56.4 on the four benchmarks…
The Iris team released the Iris-mini and Iris-pro models, as well as the SFT-RL c… on September 7, 2026.
Coverage · reports per dayLANGUAGE SPLIT
Entity relations
Integrated timelineUNIFIED TIMELINE
2026-09-07
The Iris team released the search proxy model.
The Iris team released Iris-mini and Iris-pro, trained on models with sizes of 35B-A3B and 397B-A17B respectively, and proposed the SFT-RL climbing process.
Hugging Face has released two search proxy models, Iris-mini and Iris-pro, trained respectively on 35B-A3B and 397B-A17B datasets. This study proposes the “SFT-RL climbing” process: tasks are constructed reversely from web hyperlink structures, problems are filtered and optimized through reinforcement learning, and the most difficult and efficient trajectories are returned to the supervised fine-tuning phase. With context management enabled during inference, Iris-pro achieved scores of 88.6/85.1/92.9/56.4 on the BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE benchmarks, making it the strongest open-source search proxy in its class in terms of parameter size. The authors plan to release the complete data construction, training, and evaluation recipes along with the model weights.
The Iris team has released two search agents, Iris-mini and Iris-pro, trained on datasets of 35B-A3B and 397B-A17B respectively. The study proposes the SFT-RL climbing process: tasks are constructed backward from the web hyperlink structure, data is filtered at the trajectory and round levels through supervised fine-tuning (SFT), followed by reinforcement learning (RL) optimization using a reward classifier and an observation summarizer within the training cluster, and long-range reasoning is interrupted at the request layer to retain promised prefixes. Evaluations included BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE benchmarks, with all results coming from a single ReAct agent and context management enabled when reasoning was active. With this feature enabled, Iris-pro achieved scores of 88.6/85.1/92.9/56.4 on the four benchmarks, making it the most performant open-source search agent in terms of respective parameter sizes. The team plans to release model weights and the complete…