Iris Search Agent: Open Source Context Management Beats Parameter Scaling
The Open Source Search Agent Revolution
AllSpark Research has released Iris, an open-source search agent that is changing the competitive landscape. Two models are available:
| Model | Total Params | Active Params | BrowseComp |
|---|---|---|---|
| Iris-mini | 35B | 3B | 82.2 |
| Iris-pro | 397B | 17B | 88.6 |
Both use Mixture-of-Experts architecture with 256K context windows. Weights are released under Apache 2.0.
Why Search Agents Are Different
Search agents differ from regular chat models because they must:
- Decide what to search — autonomously choose queries
- Evaluate results — determine if more searching is needed
- Know when to stop — recognize when evidence is sufficient
For two years, this capability was dominated by closed-source products. Open source had little to compete with.
Benchmark Results
Four benchmarks, impressive results:
| Benchmark | Iris-mini | Iris-pro |
|---|---|---|
| BrowseComp | 82.2 | 88.6 |
| BrowseComp-ZH | 84.8 | 85.1 |
| DeepSearchQA | 86.9 | 92.9 |
| HLE | 52.3 | 56.4 |
The Surprising Chinese Result
On BrowseComp-ZH (Chinese), the 35B model scores 84.8 and the 397B model scores 85.1 — a difference of only 0.3 points. The same models on English BrowseComp show a 6.4 point gap.
Key insight: For Chinese tasks, parameter scale barely matters. Context management is the real differentiator.
Context Management Strategy
The paper compares three inference-time context strategies:
- No management — baseline
- Discard-all — clear tool call history when context exceeds threshold, restart from original question
- Retry — summarize eliminated clues before continuing
Result: Both models improved with context management, and Iris-mini improved MORE than Iris-pro.
For small models, managing the ever-growing search history is more effective than adding parameters.
Training Data: Reverse-Engineered from Hyperlinks
Multi-hop search questions are hard to create. Writing a question that requires crossing 3-4 webpages is expensive, and models can cheat with string matching.
Iris takes a different approach:
- Pick a seed page from web corpus hyperlink structure
- Follow outgoing links to build an entity graph
- Create multi-hop chains on the graph
- Rewrite questions to ensure NO clue can be solved by literal matching
SFT-RL Climbing
Training uses a process called SFT-RL climbing: supervised fine-tuning and reinforcement learning alternate. The RL phase runs on real search, not offline snapshots. The reward model and observation summarization modules are deployed on the training cluster.
"Our policy is optimized by RL against live search."
This means every training step makes real network requests — significant engineering cost.
Why This Matters
Open source has caught up to closed source in general chat and code. Search has been the laggard.
The reason is clear: search agent capability is half model, half tool orchestration framework. Closed-source products tune both; open source often releases only the model.
Iris releases model weights, context management strategy, and data construction method — the complete package.
Community & Availability
- Hugging Face: Models available for download
- GitHub: Training pipeline code marked "coming soon"
- License: Apache 2.0 — commercial use allowed
- Reference: arXiv 2609.04304
For anyone wanting to run deep search, the 35B/3B spec is a sweet spot — runs on a single machine with the paper's context management.
Conclusion
Iris proves that open source finally has a serious competitor in search. The combination of model weights, context strategy, and data methodology closes a gap that has persisted for two years.
Tags: AI, SearchAgent, OpenSource, LLM, Agent
Top comments (0)