openJiuwen introduces X-Router for adaptive model selection in AI agents

openJiuwen introduces X-Router for adaptive model selection in AI agents

openJiuwen, an open-source AI agent platform jointly built by Huawei teams and external developers, has introduced X-Router, a model-routing algorithm within a modular routing framework. The announcement describes Ascend affinity. X-Router uses an in-process Qwen3-0.6B classifier to assign requests to five complexity tiers and select among local and cloud models. Routing considers recent conversation history, tool progress, model capabilities, cache affinity, load and availability. Execution feedback feeds a contextual bandit that can adjust subsequent selections.

The release addresses a practical constraint in the agent market: repeatedly assigning every step to a powerful model can make routine work unnecessarily expensive. It places more of the cost-quality tradeoff in the orchestration layer, where task history and operating conditions inform model selection. In a team-run evaluation covering all 147 PinchBench tasks across 11 categories, static X-Router achieved a 66.3% score at $8.25 in model-call costs. The source reports a 44.6% cost reduction with nearly unchanged scores against an all-Kimi-K2-Thinking baseline. These results support a bounded efficiency claim; they do not establish equivalent performance across production workloads or quantify Ascend-specific gains.

For builders, the concrete implication is to evaluate routing against their own task mix, measuring completion quality, latency and total cost together. The framework separates decision logic, external state and host execution, making algorithm replacement possible without changing the host or state layer. Its default retains the original selection when feedback is sparse or offers no clear advantage. That makes reliable outcome feedback central to testing whether adaptive routing improves on static selection.

#openJiuwen #AIRouting #AIAgents #Inference #HuaweiAscend