Deep search tasks force a model to coordinate web search, page extraction, stateful tool use and stepwise verification — not just to generate fluent text. XYZ-Aquila-pro embodies a pragmatic design choice: a large, open-weight checkpoint post-trained through an AI4AI pipeline to behave as a "thinking model" that emits Qwen-style reasoning traces and structured tool calls while relying on an external agent harness for concrete IO and orchestration.
Key Capabilities
- Agentic web search and browsing: produces search-focused reasoning and tool-call outputs so the surrounding harness can run targeted queries and follow-up inspections rather than relying on monolithic in-context scraping.
- Multi-source evidence aggregation and source verification: organizes and weights evidence across pages, which reduces brittle single-snippet answers and helps surface corroborated claims for downstream decision steps.
- Long-horizon planning and failure recovery: plans multi-step procedures and suggests scoped interventions when environment interactions fail, enabling iterative retrieval+execution loops instead of one-shot responses.
- Qwen-compatible tool-call formats: emits structured tool calls compatible with Qwen-style parsers, so it can integrate with OpenAI-compatible endpoints or the provided SGLang/AxisAgentic deployment stack.
Who it's for and tradeoffs
Great fit if you need an open-weight model to run research or production agents that orchestrate web search, programmatic scraping and Python evaluation under a replayable, tool-oriented harness. It is suited for teams that can provision multi-GPU deployments and supply an external agent runtime (AxisAgentic or similar). Look elsewhere if you need a turnkey hosted service, a very small on-device model, or guaranteed closed-source commercial SLAs — the checkpoint alone is not a complete agent and depends on the surrounding harness and tool implementations.
Where it fits
Positioned between smaller "agentic" mini-checkpoints (which prioritize footprint) and large closed-source search agents, this model targets sub-400B open-weight usage with competitive benchmarks in agentic search. The project emphasizes reproducible evaluation (held-out external benchmarks) and an iterative AI4AI development loop rather than one-off finetuning.
How it was developed (brief)
The checkpoint was produced by post-training Qwen3.5-397B-A17B under a bounded-exploration AI4AI pipeline: humans set goals, constraints and acceptance policies while automated agents diagnose failures and propose targeted data or training interventions. The release is intended to be used with the AxisAgentic harness and SGLang tooling for web_search / scrape / python integrations rather than as a standalone end-user product.
