PPIO正式发布“Fusion融合模型”：用十分之一的价格超越顶级模型的智商
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR — PPIO launched "Fusion融合模型," an intelligent model-gateway feature that fans a single request out to multiple "advisor" models in parallel, then has an aggregator model reconcile and merge their answers, claiming flagship-tier quality at mid-tier cost. It matters because it locates intelligence gains at the orchestration/call layer rather than in base-model parameters.
- Four-stage pipeline: request fan-out → parallel independent answers → "thinking orchestration" (extract consensus, flag disagreements, drop errors, compress context) → aggregator model writes the final response.
- Claimed DRACO deep-research benchmark result: Kimi K3 + GLM 5.2 + MiniMax M3 as advisors with DeepSeek V4 Flash as aggregator scored 57.34 vs. Claude Fable 5's 55.14, at ¥57.59 vs. ¥566 (~1/10 cost). None of the three advisors individually leads its category.
- Engineering claims: latency bounded by the slowest model (not linear in model count), graceful skip on model failure, reuse of advisor outputs within an agent turn to limit token spend, plus per-call traceability of models/tokens/latency/cost.
- Positioning: framed by PPIO's "Agent productivity = token intelligence density × agent loop duration" thesis, backed by scale claims of >1.2T daily tokens as of June 2026. Note this is a vendor-supplied piece republished by 量子位, so benchmark numbers are self-reported.
Sources (1)
PPIO正式发布“Fusion融合模型”：用十分之一的价格超越顶级模型的智商
TL;DR — PPIO launched "Fusion融合模型," an intelligent model-gateway feature that fans a single request out to multiple "advisor" models in parallel, then has an aggregator model reconcile and merge their answers, claiming flagship-tier quality at mid-tier cost. It matters because it locates intelligence gains at the orchestration/call layer rather than in base-model parameters.
- Four-stage pipeline: request fan-out → parallel independent answers → "thinking orchestration" (extract consensus, flag disagreements, drop errors, compress context) → aggregator model writes the final response.
- Claimed DRACO deep-research benchmark result: Kimi K3 + GLM 5.2 + MiniMax M3 as advisors with DeepSeek V4 Flash as aggregator scored 57.34 vs. Claude Fable 5's 55.14, at ¥57.59 vs. ¥566 (~1/10 cost). None of the three advisors individually leads its category.
- Engineering claims: latency bounded by the slowest model (not linear in model count), graceful skip on model failure, reuse of advisor outputs within an agent turn to limit token spend, plus per-call traceability of models/tokens/latency/cost.
- Positioning: framed by PPIO's "Agent productivity = token intelligence density × agent loop duration" thesis, backed by scale claims of >1.2T daily tokens as of June 2026. Note this is a vendor-supplied piece republished by 量子位, so benchmark numbers are self-reported.