NJU-spinoff SinapisAI's model router Mio tops 9 benchmarks at lowest cost
机器之心 · wechat · 2026-10-16
SinapisAI, spun out of Nanjing University's LAMDA lab, released a technical report on "Collective Model Intelligence": build quantified capability profiles for a universe of models, then let the Mio router select and coordinate them by task and budget. Its platform already lists 484 models.
Results: Mio 1.0 Lite scored the highest average (77.15/100) across 9 public benchmarks at the lowest average cost (0.06 yuan per request) — 6.35 points above GPT6.1 Sol at 94% lower cost, taking first in function calling, agent execution, GPQA and instruction following. Caveat: cost-no-object flagships still win on HLE.
Architecture: ModelEval builds cheap capability profiles via query-level modeling and cross-domain prediction; Mio maintains a dynamic candidate pool that keeps specialist models covering weaknesses; Mos picks execution endpoints by health, price and latency, with a model distribution network speeding file reads up to 23.3x. The work extends Zhou Zhihua's learnware paradigm; the company raised near-100M-yuan and plays the same "sell shovels" lane as OpenRouter (Stripe) and HuggingFace (Nvidia).
More from Companies & People
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Google's Gemini pretraining team is hiring research engineers in London and MTV — zacharynado · 2026-10-23
- After data and compute, the next picks-and-shovels are talent and physical infrastructure — philhchen · 2026-10-23
- Critics slam Anthropic for rarely defending its work against criticism on Twitter — DKokotajlo · 2026-10-23
- Amazon blocks AI shopping agents while Walmart opens its site to them with spend controls — venturetwins · 2026-10-23
- Commenters speculate Skydance could wire Grok into its entire Hollywood production pipeline — MickeySteamboat · 2026-10-23