Monday, September 14, 2026

What happened: The AllSpark team released Iris-mini and Iris-pro, two open-source search agent models built on Qwen that lead benchmarks in their size classes, along with a complete training recipe and evaluation harness.

Key details:

• Iris-mini has 35 billion parameters (built on Qwen3.6-35B-A3B); Iris-pro has 397 billion parameters (built on Qwen3.5-397B-A17B)
• Both feature a 256,000-token context window
• Training pipeline reverse-engineers tasks backward from web page link structures to create multi-step questions requiring reasoning rather than simple text search
• Two-stage filtering weeds out bad training data: first checking for correctness and repetition loops, second using a step-by-step review by a judge model
• Iris-mini scores 82.2, 84.8, 86.9, and 52.3 on BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam respectively; Iris-pro reaches 88.6, 85.1, 92.9, and 56.4
• Iris-mini leads three of four benchmarks among open-weight models in its size class
• Unexpected finding: training data and models improved performance on tasks they were never trained for, including general tool use and office work
• The Iris Harness is available on Hugging Face with benchmarks, tools, and context management strategies; data construction and training pipelines to be released later

Why it matters: These models demonstrate that open-weight search agents can achieve competitive performance with frontier systems while being transparent and reproducible. The discovery that search training generalizes to general tool use suggests search may be a foundational skill for agentic AI.

Practical takeaway: Developers looking to build search-capable agents now have high-performing open-weight alternatives; the released harness and benchmarks make evaluation and deployment straightforward.

What happened: ElevenLabs released Music v2.5 for its AI music generator, showing improved audio quality and broader genre support, with the company emphasizing use of only licensed training data.

Key details:

• In a blind test with 47,885 comparison pairs, listeners preferred v2.5 most of the time, especially for R&B, Soul, Hip-Hop, Rock, and orchestral music
• Free tier includes five lossless downloads per day; Pro tier offers 400 per month
• Users retain rights to their generated tracks; commercial use permitted depending on industry and purpose, though free tier requires attribution
• Downloads of tracks based on other artists' songs are blocked, and imitating existing musicians is not allowed
• The company's licensing deal with Universal Music Group applies only to future separate products, not Music v2.5
• v2 remains accessible alongside v2.5

Why it matters: This positions ElevenLabs' approach favorably against competitors like Suno, which faced lawsuits for training on copyrighted content without rights holders' permission. The focus on licensed data addresses growing copyright concerns in AI music generation while improving output quality.

Practical takeaway: If copyright compliance matters for your music generation use case, ElevenLabs' emphasis on licensed training is worth verifying, though the company should clarify its full data sourcing for complete transparency.