Why this matters
Agentic search—where an LLM must plan multi-step web investigations, verify sources, and recover from failed environment interactions—exposes gaps in models that only generate text. This release supplies a post-trained checkpoint that embeds a reasoning and tool-call surface designed to plug into a surrounding agent harness, enabling reproducible search workflows and stateful tool use at large context lengths.
Key Capabilities
- Qwen-compatible thinking/checkpoint: post-trained from Qwen3.6-35B-A3B to preserve Qwen-style reasoning and function-call formats so it interoperates with existing Qwen tool parsers. This means you can use the model with Qwen-aware endpoints and tooling with minimal adapter work.
- Agentic search skills: optimized for long-horizon planning, focused web browsing in English and Chinese, multi-source evidence aggregation, and explicit source verification steps rather than single-shot retrieval. The checkpoint is tuned for iterative, tool-driven research workflows.
- Harness-first deployment: the model is released as a “thinking model” rather than a full agent; concrete tool implementations, replayable context management, and evaluation pipelines are provided by the open-source AxisAgentic harness and companion data (XYZ-Aquila-SFT), not the checkpoint alone.
- High-context readiness and deployment notes: evaluation and examples assume a large context window (up to 256K) and common deployment with SGLang. Practical deployment examples show tensor-parallel setups and function-calling interfaces for controlled tool usage.
Who it fits (and tradeoffs)
Great fit if you need an open-weight model tuned specifically for structured, tool-driven web research workflows, bilingual browsing (EN/ZH), and reproducible agent evaluation with an external harness. Look elsewhere if you need an out-of-the-box conversational product, a plug-and-play multimodal model, or a tiny footprint: the intended use assumes integration with the AxisAgentic/sglang tooling, increased memory (large context/tensor parallelism), and operational scaffolding for safe tool use.
Where it fits
Positioned for AI teams building reproducible search agents and benchmarked agentic workflows: not a complete application, but a drop-in checkpoint to be combined with the AxisAgentic harness, search/scrape/python tool implementations, and the provided SFT data for agentic evaluation and deployment.
