AllSpark releases open-weight search agents Iris-mini (35B) and Iris-pro (397B) with training recipe and benchmark results

Chinese lab AllSpark released two open-weight search agent models: Iris-mini (35B parameters, built on Qwen3.6-35B-A3B) and Iris-pro (397B parameters, built on Qwen3.5-397B-A17B), both with 256,000-token context windows. Alongside the weights on Hugging Face and code on GitHub, AllSpark published a paper describing its training recipe and an evaluation harness, with results reported on four benchmarks: BrowseComp, BrowseComp-ZH, DeepSearchQA and Humanity's Last Exam. The data-construction and training pipelines themselves are not yet released; AllSpark says those are coming later.

Developers and researchers building search or research agents now have two more open-weight base models to test, with a documented harness for agent loops, tools and context management they can inspect or copy. That is genuinely useful and immediately actionable, since the weights and code are real and downloadable today. But both models are fine-tunes of existing Qwen bases rather than new foundation models, so the impact is incremental (better tooling and technique on top of known architectures) rather than a shift in who controls the underlying compute or model layer.

The headline and the "strongest in their class" claim are AllSpark's own self-reported superlative, relayed by The Decoder without independent verification or benchmark links in the coverage. The claim that search skills "generalise unexpectedly" to general tool use and office work, and that search may be a "foundational" capability, is speculative framing from the lab's paper, not a demonstrated or falsifiable result. Missing from the coverage: any comparison methodology, third-party evaluation, or explanation of what happens once the promised training pipeline is (or isn't) released.

1. 01Whether AllSpark releases the promised data-construction and training pipelines, and whether independent developers can reproduce the claimed benchmark scores using them
2. 02Third-party or independent benchmark runs of Iris-mini and Iris-pro on BrowseComp, BrowseComp-ZH, DeepSearchQA and Humanity's Last Exam, since current results are lab-reported only
3. 03Uptake signals on Hugging Face and GitHub (downloads, forks, derivative fine-tunes) over the next 4-8 weeks to see if this becomes a reference model or is quickly superseded