AllSpark ran each Iris benchmark twice to separate the model from its scaffolding

Context management is the part worth reading first. During a long research session the agent's context fills up before it has resolved all of its sub-questions ^[12]. One common fix is to discard the conversation history. That extends the run artificially and says little about the model's actual quality ^[12]. AllSpark's argument is that this runtime handling often makes a bigger difference on common benchmarks than the reported gaps between systems ^[11]. So each benchmark was scored under both conditions with the same tools, the same context limits and the same judge model on both sides ^[13]. A number published only with management on cannot be split into what came from the model and what came from the scaffolding ^[14].

Iris-pro carries about 11 times Iris-mini's parameter count ^[2][1]. With context management on, that buys 6.4 points on BrowseComp, 6.0 on DeepSearchQA, 4.1 on Humanity's Last Exam and 0.3 on BrowseComp-ZH ^[2]. Iris-mini's lead over XYZ-Aquila-mini on BrowseComp puts that model at 78.8 ^[3]. AllSpark also reports that context management helps the smaller model considerably more than the larger one ^[22].

The reported figures come from one agent, with no helper agents and no verification pass at the end ^[15]. That is closer to what a team would actually deploy than an ensemble score. For the BrowseComp number to say anything about your traffic, your questions have to look like BrowseComp's: rare facts recovered from indirect clues ^[16]. DeepSearchQA scores the completeness of the evidence retrieved, which is a different job ^[16]. The paper claims the training data and models also improved performance on tasks they were never trained for, including general tool use and office work ^[21].

The dataset construction is the most reusable piece. The pipeline starts from a seed page and its outgoing links, builds a graph of terms and relationships, and generates a multi-step question whose answer requires chaining several connected steps ^[5]. Every term except the final answer is replaced with a paraphrase, so no clue survives a plain text search ^[6]. A question enters the set only if a reference model fails it without tools and solves it with the right sources ^[7]. Teacher paths are then filtered twice: once over the whole path for correctness, repetition loops and search depth, then step by step by a judge model whose criteria were derived from the data rather than written by hand, according to the paper ^[8].

Reinforcement learning runs against a live web search, with the judge model and the result summaries hosted inside the training cluster on the team's own large Qwen model, so training does not depend on external services ^[9]. Anyone who has watched an RL run stall on somebody else's rate limit will recognise that choice. Supervised fine-tuning and RL alternate in what the authors call "SFT-RL climbing", with the hardest solved tasks and the most efficient paths from each round feeding the next cycle ^[10].

The account does not name a license for the weights or price the hardware a 397-billion-parameter model at 256,000 tokens of context needs ^[3]. The strongest-in-class claim is the team's own, and it is scoped to open-weight search agents ^[4]. The only comparison with closed systems is the authors' line that Iris-pro sometimes approaches systems that need far more compute ^[20]. On established benchmarks, according to the-decoder's account, leading systems of this kind mostly use the web to confirm knowledge they already picked up during training ^[23].