Skip to main content
Back to News
OpenSquilla releases v0.5.0 Preview with multi-model integration, top DRACO benchmarks
Product
2 min read

OpenSquilla releases v0.5.0 Preview with multi-model integration, top DRACO benchmarks

The AMW Read

Novelty=2: OpenSquilla is not a canonical case-study company but the multi-model routing pattern is a meaningful advance not yet captured in our substrate. Significance=2: The cost-performance benchmark result updates the debate on Chinese model parity and agent architecture, affecting the segment.
NoveltySignificance
AI Agents · Player MapAI Agents · Recurring Patterns

OpenSquilla releases v0.5.0 Preview with multi-model integration, top DRACO benchmarks

OpenSquilla, the open-source AI agent project developed by TokenRhythm (基元律动), released v0.5.0 Preview 1 featuring a multi-model integration system called 'agentic routing.' The routing layer orchestrates four Chinese foundation models—DeepSeek v4, GLM-5.2, Kimi K2.7, and Qwen3.7—as a parallel proposal ensemble, with a separate model aggregating outputs into a single result. No overseas flagship models are used. On the DRACO deep-research benchmark, the OpenSquilla ensemble took first place in both cost categories: on Brave Search it scored 64.09 (beating Opus 4.8 by 8.42% and GPT-5.5 by 20.27%) at an average cost of just $0.12, roughly 92% cheaper than Opus and 86% cheaper than GPT-5.5. On DuckDuckGo, it scored 60.85, essentially matching Anthropic's latest flagship Fable 5 at 59.80 while costing one-third as much ($0.39 vs. $1.21).

Why this matters: OpenSquilla's results embody the 'context-engineering moat' pattern—the thesis that orchestration and routing architecture, not individual model capability, increasingly determines real-world agent performance. By demonstrating that a heterogeneous ensemble of Chinese models can match or exceed frontier Western models at 8-33% of the cost, the project updates the open debate on whether the foundation-model gap between Chinese and Western labs has narrowed. The findings suggest that for agentic workloads, the harness layer may matter more than any single base model, potentially reshaping how enterprises evaluate build-vs-buy decisions for agent infrastructure.

From an analyst perspective, OpenSquilla's trajectory—from smart routing in v0.1.0 to self-organizing skills in v0.3.0, verifiable coding in v0.4.0, and now multi-model consensus—tracks a deliberate capital-compression play. The project is still early-stage (post-seed, ~$100M valuation) but its public benchmark positioning directly challenges the hyperscaler-distribution moat by showing that open-source routing can undercut proprietary models on cost-performance. The key risk, as noted in the DRACO methodology, is that LLM-judged benchmarks may not perfectly replicate real enterprise workflows. However, the direction of travel is clear: the battle for agent performance is shifting from model size to routing intelligence.

#OpenSquilla#multi-model routing#DRACO benchmark#cost-performance efficiency#AI agent infrastructure#context-engineering moat

How This Connects

Based on AI Agents · Player Map

  1. 2d agoAnthropic has released Claude Cowork for mobile and web, extending its enterprise AI agent beyond th...Anthropic
  2. 3d agoOpenSquilla releases v0.5.0 Preview with multi-model integration, top DRACO benchmarks · THIS ARTICLE
  3. 1w agoDatabricks Launches Omnigent, an Open-Source Platform for Multi-Agent OrchestrationDatabricks
  4. 1w agoMicrosoft introduces Agentic Resource Discovery specification for AI agents, MCP servers, and API workflows.
  5. 1mo agoUniPat AI releases SaaS-Bench, Claude Opus 4.7 passes only 3.8% of 106 real-office tasks, breaking the illusion of full office automation.

More news from OpenSquilla

Stay updated with the latest news and announcements from OpenSquilla.

View all OpenSquilla news

Discover AI Startups

Explore 2,000+ AI companies with VC-grade analysis, funding data, and investment insights.

Explore Dashboard