en.Wedoany.com Reported - The open-source AI Agent project OpenSquilla has released version 0.5.0 Preview 1, with the core update being "multi-model integration collaboration." This solution organizes four domestic models—DeepSeek v4, GLM-5.2, Kimi K2.7, and Qwen3.7—into a collaborative team for parallel proposals at the Harness layer, with a single model aggregating the final output. The lineup does not include any overseas flagship models.
Released alongside the Preview version, the technical report "Agentic Routing" explains how this harness's native routing converts daily agent traffic into a self-evolving data flywheel. The official version was subsequently released.
The latest DRACO deep research leaderboard compares the average scores and average costs of various solutions grouped by search engine. OpenSquilla's integrated solution ranks first in both groups. In the Brave Search group, the average score is 64.09, higher than standalone Opus 4.8 (59.11, +8.42%) and GPT-5.5 (53.28, +20.27%); the average task cost is $0.12, approximately 92% and 86% lower respectively, making it the only solution in this group to achieve both the "highest score" and "lowest cost" markers. In the DuckDuckGo group, the average score is 60.85, slightly higher than Anthropic's latest flagship Fable 5 at 59.80, with scores essentially on par, but the cost is about one-third ($0.39 vs. $1.21); Fable 5's results in the Brave group are still being processed.
The mechanism of this solution is "diversity sampling + consensus aggregation": multiple models independently complete search and reasoning, complementing each other to compensate for the inherent shortcomings of a single model, such as missing information sources, miscalculating values, or failing to cover all constraints. The team stated that this is not about switching to a stronger model, but about adopting a better organizational approach. This result points to a judgment: while domestic base models individually still lag behind overseas flagships, when properly organized at the Harness layer, a mix of domestic models can achieve higher and more stable scores on real tasks, and even match or surpass the latest generation of flagships at a fraction of the cost.
OpenSquilla is developed by TokenRhythm, positioning itself with a dual-track focus on Harness and model optimization, with the product philosophy of "enhancing Agent intelligence per unit cost." Version evolution revolves around "spending less, delivering real results": v0.1.0 introduced intelligent routing, automatically selecting models based on task difficulty; v0.2.0 launched one-click migration, supporting low-cost switching from other Agent frameworks; v0.3.0 released the MetaSkill self-organizing skill protocol; v0.4.0 brought verifiable coding and the first signed desktop version; leading to the multi-model integration in this v0.5.0 Preview. According to public reports, the company completed its first round of financing shortly after its establishment, with a valuation of $100 million.
