openJiuwen debuts X-Router self-evolving model routing with Ascend affinity, cutting agent token consumption by over 50%

On October 2 the open-source AI agent platform openJiuwen released X-Router, a self-evolving model routing technology that inserts a request-level decision layer between the agent and its pool of candidate models. openJiuwen is built jointly by teams including Huawei's 2012 Labs, Huawei Cloud, and its device, computing and computing-power units together with universities and outside developers, and X-Router is the first systematic public disclosure of that routing capability.

As the report frames it, most AI applications still rely on one model for everything: a request to translate a sentence and a request that requires writing code and debugging it three times both land on the same hundred-billion-parameter model. Once an agent holds local models, cloud models and services from several vendors at the same time, a hard-coded routing table cannot keep up with model iterations and cannot balance cost against quality.

openJiuwen splits routing into three layers. The first builds capability profiles for each model, drawing them offline and refreshing them online across task types, difficulty levels, latency, power and cost, with profile updates landing in under a minute when a model version or its behaviour changes. The second handles the decision itself, taking in not just the complexity of the current request but the whole multi-turn trajectory and the state of the system right now. The third makes the policy evolve as real feedback accumulates.

The decision layer folds in signals that pure algorithms usually miss. It reads KV-cache affinity to judge which model already holds a warm cache for this context, checks real-time load so a request is not pushed onto a busier model, and writes recently timed-out or rate-limited models into state as exclusions for the next round. The algorithm itself is a pure function: the same request and the same state snapshot must produce the same decision, and all cross-request memory lives in a separate state layer, so losing state only downgrades to "cold routing" instead of failing the request.

Above model selection, the router can orchestrate multi-model collaboration. When one model is not enough for a trustworthy answer, the routing layer decides whether to launch several models this round and which ones, and WorkSwarm's MoA mechanism handles semantic de-duplication, quality filtering, conflict resolution and compression. The team describes the granularity shifting from "choosing a model" to "orchestrating the whole execution strategy", placing models, tools and sub-agents inside one scheduling framework.

The measurements come from x-router plugged into WorkSwarm and evaluated on the full 147-task PinchBench set, spanning 11 categories including log analysis, data analysis, coding and research, with a locally deployed Qwen3-0.6B classifier running in-process. Sending everything to the cloud model Kimi-K2-Thinking scored 71.36% at a cost of $11.11; x-router's static routing scored 66.3% at $8.25, cutting cost by 25.7%; switching on the Bandit self-evolution mode pushed cost down to $6.16 while the score recovered to 70.70% - roughly 44.6% cheaper than the baseline.

A second test on Terminal-Bench/LLMRouterBench used a three-tier pool of Opus4.8, Qwen3.5-122B and Qwen3.5-35B. In cost-focused mode it cut cost by 51.4% while reaching 95.2% of Opus's success rate; in quality-focused mode it cut cost by 15.7% while reaching 98.6%. Both benchmarks point the same way: near-top-tier quality at roughly half the cost, with how much quality to trade for how much saving left as a user-configurable choice.

For agents, routing is turning from a nice-to-have optimization into a precondition for scale. Unlike a chatbot, an agent works continuously across many turns and tools, so token consumption multiplies rather than growing linearly; openJiuwen says the kernel aims for generality rather than a single-point optimum, with one framework configurable into edge, cloud, on-premise and hybrid deployments. The core is written in Rust with a Python layer via PyO3, and the code is open-sourced in the openJiuwen/model-router repository on gitcode.