PPIO Unveils Fusion Model: Top-Tier Intelligence at One-Tenth the Cost
Anthropic
Moonshot AI
MiniMax
DeepSeek
PPIO has launched its Fusion Model, a new feature of its intelligent model gateway that orchestrates multiple AI models to improve answer quality and reduce costs. In benchmarks, it outperforms Claude Fable 5 with a score of 57.34 versus 55.14, at roughly one-tenth the cost. The Fusion Model demonstrates that intelligence can be enhanced at the orchestration layer, not just through larger base models.
On August 6, PPIO officially released the Fusion Model, a new intelligent model gateway feature. Unlike traditional API gateways that simply forward requests, Fusion distributes complex tasks to multiple specialized models in parallel, then uses orchestration and cross-validation with a main model to produce a final answer. In the DRACO deep research benchmark, PPIO combined three open-source models to score 57.34, surpassing Claude Fable 5's 55.14, with a total cost of ¥57.59 compared to ¥566, about one-tenth the cost. The Fusion process involves four steps: request distribution, parallel answering, orchestration (extracting consensus, flagging divergences, and compressing context), and final aggregation. This approach addresses issues like model specialization and hallucination by providing multiple reasoning paths. In tests, PPIO used Kimi K3, GLM 5.2, and MiniMax M3 as advisors with DeepSeek V4 Flash as aggregator. PPIO claims that as of June 2026, its daily token consumption exceeds 1.2 trillion tokens, ranking first among independent AI cloud service providers in China. The company was selected for the CAICT enterprise token service performance baseline.
- Abbreviations
- API = Application Programming Interface — интерфейс программирования приложений
- WAIC = World Artificial Intelligence Conference — Всемирная конференция по искусственному интеллекту
- CAICT = China Academy of Information and Communications Technology — Китайская академия информационных и коммуникационных технологий
- TPS = Transactions Per Second — транзакций в секунду
- TTFT = Time To First Token — время до первого токена
Source: QbitAI 量子位 —
original