The Pulse
AllSpark Releases Open-Weight Iris Search Agents
AllSpark Research’s Iris-mini model card details a 35B-parameter open-weight search agent, while the accompanying arXiv paper describes its training and evaluation methods.
AI.info Team ·
AllSpark Research has released Iris-mini, an open-weight search agent designed for long, multi-step web research. Its Hugging Face model card identifies a larger sibling, Iris-pro, and links to an evaluation harness and project repository. A paper submitted to arXiv on September 3, 2026, “Iris: Climbing to the Search Frontier”, describes the models’ data construction, training procedure and evaluation setup.
Iris-mini is post-trained from Qwen3.6-35B-A3B. The model card lists 35 billion total parameters and 3 billion active parameters, with 256 experts and eight active experts. It has a 256,000-token context window, uses bfloat16 precision and is available under the Apache 2.0 license. Iris-pro is listed as a 397B-A17B model.
Two Iris models target long web searches
The model card describes search as a loop in which an agent decides what to look up, reads returned material, determines whether to continue and stops when it has enough evidence. It says Iris is trained through alternating rounds of supervised fine-tuning and reinforcement learning.
The paper provides the additional detail behind that description. Its authors say they construct multi-hop questions from the hyperlink structure of a web corpus, rewrite non-answer entities into descriptive references and admit questions that a reference model fails to answer without supporting evidence. They then create and filter search trajectories before supervised fine-tuning, and optimize the policy with reinforcement learning against live search.
The paper calls the alternating procedure SFT-RL climbing. Under that process, hard solved trajectories and efficient rollouts from a reinforcement-learning round feed into a subsequent supervised-training pass.
Ziyuan Liu, the paper’s first-listed author, and eight co-authors describe the evaluation setup this way:
“All results come from a single ReAct agent, with no sub-agents and no test-time verification.”
That setup matters because long search tasks can produce materially different results depending on the handling of accumulated context and failed attempts.
Iris-mini reaches 82.2 on BrowseComp
Using the default discard-all context-management setting, Iris-mini scores 82.2 on BrowseComp, 84.8 on BrowseComp-ZH, 86.9 F1 on DeepSearchQA and 52.3 on the 2,158-question text-only subset of Humanity’s Last Exam, according to the model card. BrowseComp, BrowseComp-ZH and Humanity’s Last Exam are reported as accuracy measures.
The card compares Iris-mini with open-weight systems in the 30-billion-to-35-billion-parameter range. It lists Iris-mini above the other named systems on BrowseComp, BrowseComp-ZH and Humanity’s Last Exam. XYZ-Aquila-mini has the higher listed DeepSearchQA score, at 89.5 compared with Iris-mini’s 86.9. The card says the comparison systems’ figures come from their public reports and may use their own context-management methods.
The paper reports higher figures for Iris-pro: 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 on DeepSearchQA and 56.4 on Humanity’s Last Exam. It characterizes both Iris models as the strongest open-source search agents in their respective parameter ranges across those four benchmarks.
Context resets account for a large share of the gains
AllSpark reports Iris-mini results under several context-management settings. Without context management, the model card lists scores of 64.7 on BrowseComp, 72.3 on BrowseComp-ZH, 81.0 on DeepSearchQA and 43.2 on Humanity’s Last Exam.
Under discard-all, the conversation is reset to the opening question after the prompt crosses a threshold. That setting raises the listed BrowseComp score to 82.2. A separate retry method restarts an episode that ends without a parseable answer while carrying forward a short summary of what has been ruled out. Combining the two methods produces an 85.9 BrowseComp score, though AllSpark reports discard-all as its headline setting even where retry produces a higher result.
The release includes a local evaluation path
The Iris-mini model card provides instructions for serving the model with SGLang and evaluating it with Iris-Harness. Its example specifies four-way tensor parallelism, a 262,144-token context length, Qwen reasoning and tool-call parsers, and a local OpenAI-compatible endpoint.
The card says the model is trained to call tools through the OpenAI function-calling interface and to wrap final answers in \boxed{}. It also says prior reasoning is carried back into the conversation on each turn and that the harness configuration supports the published context-discard threshold.
Open weights, but not a turnkey search service
Iris-mini’s weights are available under Apache 2.0, with instructions for local serving. The Hugging Face page states that the model is not deployed by an Inference Provider, leaving users to run compatible infrastructure themselves. The release therefore offers researchers both the model and a documented path to test how the search loop and context-management settings affect its reported performance.