Iris Review: In-Depth Analysis of Xiaohongshu AllSpark Team's Open-Source Search Agent

1. What is Iris

Iris is a search agent open-sourced by the Xiaohongshu AllSpark team, featuring two versions: 35B (Iris-mini) and 397B (Iris-pro). It utilizes a MoE architecture, with activated parameters of 3B and 17B respectively. This model is capable of autonomously deciding search strategies, reading methods, and convergence timing within real-world network environments. By constructing training data in reverse from webpage hyperlinks, and combining SFT with reinforcement learning through the SFT–RL Climbing training method, Iris achieves state-of-the-art performance on search benchmarks such as BrowseComp, DeepSearchQA, and HLE among open-source models of similar scale. The 35B version has already approached the search capabilities of trillion-parameter models, and its search ability can be extended to tasks such as general tool calling.

Technical Positioning and Domain: Iris belongs to the search agent domain, focusing on enabling large language models to complete closed-loop tasks involving autonomous retrieval, multi-step reasoning, and evidence integration in real-world network environments. Unlike traditional RAG approaches, Iris treats the search engine as a callable tool, with the model itself determining the search strategy, depth of reading, and answer convergence conditions, thus achieving a "search, think, and respond" reasoning paradigm.

Development Background: This model was developed by the Xiaohongshu AllSpark team, which has established a complete technical system in data construction and training engineering. The Iris project was open-sourced in 2026, releasing model weights, inference and evaluation code, data construction methods, and training recipes, with the aim of promoting the transparency of search agent technology.

Core Value: Iris addresses three major pain points of existing search models in real-world networks: "single search strategies, weak cross-page reasoning capabilities, and reliance on manually annotated training data." By implementing a reverse question generation mechanism, it enables automated and large-scale data construction. Additionally, through the alternating SFT–RL Climbing training method, it forms an adaptive curriculum, allowing the model's capabilities to continuously improve as training progresses.

Technical Features: The core technical highlight is its "reverse" data construction logic—generating questions that require multi-step cross-page reasoning from the hyperlink graph of webpages, while removing any matching clues. Furthermore, it implements mechanisms such as fully internalized scoring and summarization, and checkpoint resuming for ultra-long trajectories, significantly enhancing training efficiency and stability.

2. Key Features

• Autonomous Web Search: The model independently decides on search keywords, analyzes the relevance of retrieved results, and determines when to continue searching or converge on an answer. This capability is trained using reinforcement learning, enabling the model to develop dynamic search strategies in real-world web environments, rather than following a fixed, predefined process.

• Multi-step Cross-page Reasoning: Supports step-by-step reasoning across multiple web pages to pinpoint a unique answer, combining factual snippets from different pages into a complete chain of evidence. Unlike shallow retrieval that relies on keyword matching, Iris can perform multi-hop reasoning by following hyperlinks, making it suitable for complex questions where information is scattered and requires logical connections.

• Multilingual Complex Retrieval: Achieves state-of-the-art performance among open-source models on both the BrowseComp and BrowseComp-ZH benchmarks, demonstrating stable handling of mixed Chinese-English corpora and cross-lingual information retrieval.

• Evidence-Coverage Based Q&A: When answering, the model actively covers the full chain of evidence rather than providing isolated conclusions. It performs exceptionally well on the DeepSearchQA benchmark, making it suitable for scenarios requiring high traceability of answers, such as in-depth research and fact-checking.

• Expert-level Problem Solving: Capable of handling highly complex problems such as HLE (Humanity's Last Exam), which require expert-level knowledge and multi-step reasoning. This indicates that its reasoning depth extends beyond common-sense questions and can be applied to specialized domains.

• Context Management Strategy: Implements context management mechanisms such as truncation and compression, supporting uninterrupted long-chain searches. This design is particularly crucial for Iris-mini, a model with fewer parameters, effectively mitigating issues of information loss and performance degradation under long context conditions.

• Cross-domain Capabilities: Without specific training, the model can handle general-purpose tool calling and office collaboration tasks, performing well on benchmarks such as BFCL, τ-bench, OfficeQA, and APEX. The "decision planning—tool calling—information integration" skills acquired during the search process are transferable across domains.

3. How to Use

1. Clone the code repository: Execute git clone https://github.com/AllSpark-Research/Iris to obtain the complete inference and evaluation code. The repository includes model loading, search tool integration, and evaluation scripts.

2. Download model weights: Download the Iris-mini (35B) or Iris-pro (397B) weights from the HuggingFace collection huggingface.co/collections/AllSpark-Research/iris. Iris-mini requires 3B activation parameters, while Iris-pro requires 17B activation parameters. Choose the appropriate version based on your inference resources.

3. Set up the runtime environment: Install the required dependencies according to the repository instructions, including inference frameworks such as vLLM or SGLang. The network environment must allow access to external search engine APIs, as internet-based retrieval is a prerequisite for the search agent to operate.

4. Load the model and integrate the search tool: Load the model using the ReAct agent format and register the search engine as a tool for the model to call. Iris's search strategy relies on real network feedback, so it is essential to ensure that the search interface is available and that the returned results maintain consistent quality.

5. Run inference: After entering a question, the model automatically executes the full process of query planning—web page retrieval—multi-step reasoning—answer convergence. It is recommended to perform initial tests with questions that have verifiable facts to observe the model's search strategy and the quality of its evidence citations.

6. Reproduce evaluations and fine-tune: The repository includes an aligned evaluation protocol that unifies tools, context, and scoring models, enabling the reproduction of benchmark results such as BrowseComp. If you need to adapt to specific business scenarios, you can adjust hyperparameters such as context management strategies (truncation/compression) and search call frequency.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For research institutions and enterprises that require full control over the internal mechanisms of the model, Iris is currently a highly open-sourced option in the field of search agents. Its data construction methods and training recipes are reproducible, making it convenient for further fine-tuning and customization in specific vertical domains. In scenarios with limited computational resources but a focus on search quality, the 35B scale of Iris-mini (activated 3B) provides high parameter efficiency. However, for products requiring composite capabilities such as general conversation and content creation, trillion-parameter general models like Kimi-K2.6 offer greater overall advantages. Nevertheless, its Agent Swarm evaluation method relies on the addition of multiple agents, and comparisons with single-model search capabilities should take protocol differences into account.

6. Editor's Summary

Iris represents a significant breakthrough in technological innovation within the field of search agents. Traditional search-augmented generation approaches typically employ a pipeline model of "retrieval—concatenation—generation," with fixed retrieval strategies and limited cross-document reasoning. Iris, however, utilizes an alternating training method of SFT–RL Climbing, enabling the model to autonomously develop search strategies within real network environments and achieve dynamic search decision-making. Its data construction mechanism for generating "sufficiently challenging yet solvable" training data from web hyperlink graphs is particularly noteworthy—this process requires no manual question creation and can scale infinitely with the size of the corpus, addressing the industry challenge of scarce training data for search agents.

In terms of practical value, Iris-mini achieves a score of 82.2 on BrowseComp with just 35B parameters, approaching the performance of trillion-parameter models, which demonstrates high parameter efficiency. For research teams and enterprises with limited computational budgets, this means they can obtain usable search agent capabilities without relying on ultra-large models. Additionally, the ability of its search functionality to extend to general tool calling makes it a promising candidate as a foundational model for agents.

In terms of target users, Iris is suitable for product developers engaged in in-depth research, search-augmented application engineers, researchers focused on LLM inference and training, and academic institutions requiring reproducible search capabilities. Its open-source scope includes weights, evaluation code, data construction, and training recipes, providing a complete reference system for secondary development and technical research.

Looking ahead, the integration of search agents with general-purpose agent capabilities is becoming increasingly evident. The direction validated by Iris—"search as a reusable atomic capability"—is expected to find application in more product forms. If the team further optimizes the model's performance in managing context within extremely long chains of reasoning and expands the range of evaluation benchmarks, the value of its open-source ecosystem will be further realized.

7. Application Scenarios

• Deep Research and Survey Reports: For scenarios involving industry analysis, competitive research, and academic literature reviews, Iris can automatically perform multi-round retrieval and cross-page reasoning, piecing together evidence scattered across different web pages into comprehensive research conclusions. It is suitable for Deep Research products and can significantly reduce the time and cost of manual information collection and integration.

• Complex Fact-Checking: In scenarios requiring source tracing such as debunking misinformation, fact-checking, and due diligence, Iris provides answers that cover the complete evidence chain. It can identify the unique answer across multiple web pages and present a traceable reasoning process, meeting the strict requirement of "must find the source."

• Bilingual Information Retrieval in Chinese and English: Iris achieves leading performance on both BrowseComp and BrowseComp-ZH, capable of handling retrieval tasks with mixed Chinese and English content. It is applicable to scenarios requiring cross-verification in Chinese and English, such as cross-border information monitoring, overseas market research, and multilingual data integration.

• General Tool Calling Agent Foundation: Iris's capabilities can be extended to general tool calling benchmarks such as BFCL and τ-bench, without the need for specialized training, to handle tasks like tool selection, parameter generation, and result parsing. It can serve as a foundational model for building General Tool Use Agent products.

• Office Collaboration Assistant: On office-related tasks such as OfficeQA and APEX, Iris can handle document Q&A and process assistance tasks without specialized training. It is well-suited for integration into office assistant products to manage complex work instructions that require retrieval capabilities.

8. FAQ

Q: What is the fundamental difference between Iris and traditional RAG (Retrieval-Augmented Generation) approaches?
A: Traditional RAG typically treats retrieval as an independent pre-processing step, performing one or more vector searches based on the user's question and then incorporating a fixed number of documents into the context. Iris, on the other hand, treats the search engine as an interactive tool, with the model autonomously deciding on search keywords, reading strategies, and when to stop, forming a dynamic closed-loop of "searching while reasoning." Its search trajectory is not a linear "question—retrieve—answer" process, but rather a multi-step decision-making process.

Q: How to choose between Iris-mini and Iris-pro?
A: Iris-mini has 35B parameters (activated 3B) and achieved a score of 82.2 on BrowseComp; Iris-pro has 397B parameters (activated 17B) and scored 88.6. If your inference resources are limited or you are sensitive to latency, Iris-mini can provide near-large-model search capabilities with less computational power. If you are pursuing the highest search quality and have ample computational resources, Iris-pro is more suitable.

Q: What hardware configuration is required to deploy Iris?
A: Iris-mini requires only 3B activated parameters and can run inference on a single consumer-grade GPU (e.g., RTX 4090 24GB). Iris-pro requires 17B activated parameters and is recommended to be deployed on multi-GPU A100/H100 setups or servers with large GPU memory. Specific configurations depend on the inference framework (vLLM/SGLang) and concurrency requirements.

Q: Can Iris work if the internet search fails?
A: No, it cannot function properly. Iris's search strategy relies on real network feedback, and if the search API is unavailable or returns low-quality results, the model will be unable to effectively perform multi-step reasoning. Before deployment, ensure that the search engine interface is stable and has sufficient calling quotas.

Q: Can the evaluation results of Iris be reproduced locally?
A: Yes. The Iris repository includes an aligned evaluation protocol, unifying the tools, context, scoring model, and context management strategies, and limiting the evaluation to a single ReAct agent. By configuring the environment according to the repository instructions and downloading the corresponding evaluation dataset, you can reproduce the performance of benchmarks like BrowseComp locally.

Q: Is the method for constructing Iris's training data publicly available?
A: Yes. The project has publicly disclosed the methodology for generating training data from web hyperlinks, including key steps such as entity graph construction, question generation, clue removal, and dual screening with a reference model. The data construction process has been automated and can scale with the growth of the corpus.

9. Project Links

• GitHub Repository: https://github.com/AllSpark-Research/Iris (includes inference, evaluation code, and training recipe documentation)
• HuggingFace Model Library: https://huggingface.co/collections/AllSpark-Research/iris (provides download of Iris-mini 35B and Iris-pro 397B model weights)
• arXiv Technical Paper: https://arxiv.org/pdf/2609.04304 (systematically explains data construction methods, SFT–RL Climbing training approach, and evaluation protocol)