SinapisAI's Mio1.0Lite tops 9 benchmarks at ~6 cents per call by coordinating cheap open models
新智元 · wechat · 2026-10-18
AI infra startup SinapisAI released a technical report introducing "Collective Model Intelligence" and Mio1.0Lite, which coordinates multiple cost-effective Chinese open models. It averaged 77.15 across 9 public benchmarks (agent automation, function calling, instruction following, science QA), ranking first among 13 systems, at 0.062 RMB per call — 94% cheaper than GPT-6.1Sol and over 60x cheaper than Claude Opus 5.5. Code and datasets are open-sourced. The system has three components: ModelEval (estimates per-domain capability profiles from only 5-40 sample questions), Mio (maintains a complementary candidate model pool and routes each task step by budget, ranking first on the 35,278-instance RobustRouteBench), and Mos (unified serving with shared caching, cutting an 18.84GB model's load time from over an hour to 3 minutes). The idea traces to Zhou Zhihua's "learnware" paradigm and is framed as a new dimension of scaling laws. The authors note results are component-level validation only, with ablations, multi-step collaboration, and live scheduling still to be evaluated.
More from Models
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23
- Marker's lite mode scores 75.1 vs 42.3 for Azure Read on OmniParseBench — VikParuchuri · 2026-10-23
- Datalab launches lite OCR at $0.50 per 1,000 pages, 20-30x cheaper than Azure, Google and AWS layout APIs — VikParuchuri · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- PHP Foundation reveals Anthropic's Claude cyber model scanned PHP, validated with OpenAI's Daybreak Red — saibharadwaj · 2026-10-23