Huawei's openJiuwen Launches X-Router Self-Evolving Model Routing Engine With Ascend Chip Affinity

Dynamic Routing for Multi-Model Agent Workloads

Huawei's open-source AI agent platform openJiuwen has launched X-Router, a model routing engine that dynamically assigns each agent request to the most appropriate model in a mixed pool of local and cloud-based options. The system addresses a growing pain point in production AI agents: as model ecosystems expand, static routing rules fail to keep pace with changing model capabilities, pricing, and workload conditions.

openJiuwen is built jointly by Huawei 2012 Laboratory, Huawei Cloud, and other internal teams alongside university and industry contributors. The platform describes itself as an open-source AI agent framework with native affinity for Huawei's Ascend chip ecosystem, positioning X-Router as infrastructure for scaling agent deployments on domestic compute hardware.

Three-Layer Architecture: Profiling, Decision, Evolution

X-Router's architecture comprises three decoupled layers. The profiling layer maintains dynamic capability profiles for each candidate model, updated within one minute when model versions or performance characteristics change. Profiles capture per-task-type accuracy, latency, power consumption, and cost rather than static scalar scores.

The decision layer evaluates each incoming request against the full conversation trajectory, KV-cache affinity across models, real-time load, and user-configured preferences such as cost-first or quality-first. It formulates model selection as a multi-objective optimization problem solved via heuristic algorithms, choosing the best-fit model rather than the most powerful one.

The evolution layer closes the feedback loop. After each model call, execution results including success/failure status, latency, cost, and quality scores are written back to an externalized state store. A Contextual Bandit algorithm retrieves similar historical experiences to refine future routing decisions. When sample sizes are insufficient or advantages unclear, the system conservatively retains its original decision rather than forcing adaptation.

Benchmark Results Across Two Evaluation Suites

The team evaluated X-Router integrated into their WorkSwarm agent framework using PinchBench's full 147-task suite spanning log analysis, data analysis, coding, and research. Using a five-tier model pool ranging from lightweight local models to cloud-based reasoning models, X-Router with static routing scored 66.3% at $8.25 total cost, compared to the all-cloud Kimi-K2-Thinking baseline of 71.36% at $11.11. Enabling Bandit self-evolution improved the score to 70.70% while further reducing cost to $6.16, representing a 44.6% overall cost reduction with only a 0.66 percentage-point quality gap.

On Terminal-Bench/LLMRouterBench with a three-model pool of Claude Opus 4.8, Qwen3.5-122B, and Qwen3.5-35B, cost-focused routing achieved 95.2% of Opus-level success rates at 51.4% lower cost. Quality-focused routing reached 98.6% of Opus success rates at 15.7% lower cost. Both benchmarks used the same evaluation models and scoring criteria across all experimental conditions.

Open-Source Release and Deployment Flexibility

X-Router's core is written in Rust with Python bindings via PyO3, and is available on both GitHub and AtomGit under the openJiuwen project. The local complexity classifier uses Qwen3-0.6B running in-process, requiring no separate service. Configuration is handled through TOML files specifying algorithm choice, model pool, and state backend parameters.

The team outlines four deployment topologies supported by the same routing logic: edge (single-process, in-memory, zero external dependencies), cloud (externalized state with global load visibility), enterprise private (local model pools with data residency), and hybrid edge-cloud (simple tasks handled locally, complex tasks escalated to cloud). Future work includes reinforcement learning-based training to convert short-term contextual bandit experience into more stable long-term routing policies.