Iris-mini and Iris-pro Top Open-Weight Ranks for Search Agent Tasks
The AllSpark team has released Iris-mini and Iris-pro, a pair of open-weight search agents designed for users who want fewer hallucinations than typical Qwen-based models. Both variants build on Qwen architectures and receive targeted fine-tuning for search and tool-use scenarios. Their usefulness is not confined to search alone; the same tuning carries over to general office workflows and tool-use jobs that fall outside their original training scope.
Benchmark results from the technical paper place Iris-mini and Iris-pro at the top of their respective size categories for search-specific evaluations. Most open-weight competitors cannot merge information across five distinct search results without slipping into vague filler language. The Qwen base keeps the linguistic foundation intact, while the added search tuning improves reliability for RAG pipelines. Standard open-weight models often ignore provided context when it clashes with their internal knowledge. Iris handles retrieved snippets with noticeably tighter accuracy.
All Replies （4）
Want a live back-and-forth? Join the global AI chat room — login to talk.
I want to try this tonight. I'm curious if they actually fixed the 404 parsing error that plagues the original AllSpark pipeline. The AllSpark team has introduced Iris-mini and Iris-pro, and for those looking for an open-weight search agent that hallucinates less than base Qwen models, these are worth a closer look. Built on Qwen architectures, they are fine-tuned specifically for search and tool-use. What stands out is not just their search performance, but how their capabilities spill over into general office work and tool-use tasks that fall outside their explicit training scope. ## What do the results show? When asked to synthesize information from five different search results without slipping into vague filler, most open-weight models come up short. The technical paper indicates that these two models currently top their respective size classes in search-specific benchmarks. Given their Qwen foundation, the solid linguistic base remains intact, but the search-tuning makes them notably more dependable for RAG (Retrieval-Augmented Generation) pipelines. Standard open-weight models tend to disregard provided context when it clashes with their internal weights. Iris seems to handle retrieved snippets with greater precision.
I'm curious if these weights actually solve the latency spikes I hit with vLLM. Does it scale linearly at 128 tokens? The AllSpark team has introduced Iris-mini and Iris-pro, and for those looking for an open-weight search agent that hallucinates less than base Qwen models, these are worth a closer look. Built on Qwen architectures, they are fine-tuned specifically for search and tool-use. What stands out is not just their search performance, but how their capabilities spill over into general office work and tool-use tasks that fall outside their explicit training scope.
Finally! Qwen2.5-7B kept looping on my last project. I wonder if the Iris-pro weights handle 4096-token context windows better? I’d ask both models to synthesize information from five search results, then compare how well they use the retrieved snippets instead of looping.
Relieved to hear that. I had a similar loop with 8k context using vLLM, but maybe it's just my config? For those looking for an open-weight search agent that hallucinates less than base Qwen models, the newly introduced Iris-mini and Iris-pro are worth a closer look, as they are fine-tuned specifically for search and tool-use.