Beyond Compute, Algorithms, Data: Collective Model Intelligence as AI's Fourth Growth Factor

AI capability gains have historically relied on scaling three pillars: compute, algorithms, and data. The article argues that models themselves are emerging as a fourth growth factor, driven by two shifts: models are becoming reusable, accumulable resources with diverse skill trees, and marginal returns from scaling single models are diminishing. The untapped potential lies in complementary collaboration among massive model populations.

Collective Model Intelligence Framework

The concept extends the 2016 "Learnware" paradigm proposed by Academician Zhou Zhihua (Nanjing University), which advocated models with specifications, shared repositories, and on-demand reuse. Collective Model Intelligence focuses on capability-level coordination: understanding and orchestrating independently developed models' complementary abilities to improve task outcomes and resource efficiency. This aligns with China's "15th Five-Year Plan" emphasis on group intelligence innovation and intensive, green compute development.

System Architecture: Four Interlocking Components

Model Eval — Capability Profiling

Full evaluation of every model on every benchmark is prohibitively expensive. Model Eval combines query-level response modeling with cross-domain prediction, using historical evaluation data and limited new observations to build capability profiles. In experiments covering 25 models, 18 benchmarks, and over 25,000 queries, the system's average score estimation error was lower than IRT and PSN-IRT baselines. A separate 35-model, 32-domain experiment validated cross-domain prediction using tested domains as anchors.

Mio — Model Selection and Collaboration

Mio integrates capability evidence with task state: it first filters candidates satisfying interface, modality, and budget constraints, then compares expected task performance, and finally picks the most cost- and latency-efficient option among similarly performing models. The candidate pool is dynamically maintained as new models join, capability evidence updates, and service conditions change. A directional replacement loss metric ensures that models with unique strengths on specific requests are retained even if their average scores are unremarkable.

Mos — Model Service Orchestration

Identical model variants across providers differ significantly in pricing, rate limits, queue latency, generation speed, and availability. Mos schedules execution endpoints based on health, capacity, cost, and estimated latency, with conditional retries and same-model endpoint failover. Its model distribution network, leveraging pre-warmed shared caches, accelerates model file reads up to 23.3× (excluding cache warm-up, parameter loading, and inference).

Model Space — Versioned Model Repository

The underlying repository assigns versioned identities to each model, enabling dynamic onboarding, updates, and retirement of 484 models (as of public beta launch), with capability profiles refreshed per version.

Benchmark Results: Mio 1.0 Lite

Mio 1.0 Lite was evaluated against 8 top-tier single models and 4 routing services on 9 public benchmarks:

Average score: 77.15/100 (highest overall)

Average cost per request: ¥0.06 (lowest overall)

Won 5 of 9 datasets outright

vs. highest-scoring competitor (GPT 6.1 Sol): +6.35 points, 94.0% cost reduction

vs. lowest-cost competitor (Qwen 3.8 Flash): +14.00 points, 16.4% further cost reduction

Per-benchmark highlights:

AutomationBench (Agent execution): 53.83% — 1st, +4.00 pp over Jev Router

BFCL V3 Base Multi-Turn (function calling): 77.50% — 1st, +11.50 pp over Opus 5.5, cost 2.7% of Opus

GPQA Diamond (scientific reasoning): 95.96% — 1st, +2.46 pp over Kimi K3

Humanity's Last Exam (tool-free reasoning): 53.38% — 2nd, -7.97 pp vs. Opus 5.5, cost 97.5% lower

AA-OmniScience (factual knowledge): 64.67% — 2nd, -0.67 pp vs. GPT 6.1 Sol, cost 85.3% lower

PinchBench (Agent execution): 96.27% — 1st, +0.23 pp over GLM 5.3

QwenClawBench (Agent execution): 69.38% — 2nd, -0.43 pp vs. Qwen 3.8 Max, cost 95.1% lower

IFBench (instruction following): 92.00% — 1st, +9.20 pp over Qwen 3.8 Max, cost 91.6% lower

MT-Bench (multi-turn chat): 9.138/10 — 3rd, -0.223 vs. GPT 6.1 Sol, cost 89.7% lower

The report acknowledges boundaries: on high-difficulty tasks like HLE, flagship models without cost constraints retain quality advantages. Mio targets a better quality-cost-coverage balance. RobustRouteBench was constructed to evaluate request-level selector robustness across 18,982 test requests in general text, multimodal, and Chinese scenarios, avoiding development-set leakage.

Commercial Context and Positioning

Recent acquisitions signal market validation: Stripe acquiring OpenRouter (model gateway/routing) and Nvidia acquiring Hugging Face (largest model repository) for $12.9B. Capital is pricing the "water sellers" connecting supply and demand. SinapisAI's stack maps to three layers: model discovery (Model Square), selection and collaboration (Mio), and capability profiling plus execution support (Model Eval, Mos, Model Space). The company raised nearly ¥100M in Series A from Nanjing Innovation Investment and Dingxin Capital. Crucially, SinapisAI does not train general-purpose models, maintaining neutrality as a scarce asset.

Limitations and Vision

The report explicitly states that system-level integrated gains and the vision of "collective intelligence evolving continuously with experience" require further verification. The envisioned recursive evolution loop uses collective capability to diagnose system weaknesses, correct mechanisms, and validate changes — each cycle aiming to improve both task performance and self-improvement capacity. This is not a claim to end Scaling Laws, but a pragmatic observation: when single-model capability growth shows diminishing marginal returns, organizing existing models' capabilities constitutes a new progress curve. CEO Sun Tengzhong analogizes the mission to building the power grid for the model era: not generating models, but ensuring every model capability is visible, schedulable, and delivered to the tasks that need it most.

Technical report: https://github.com/SinapisAI/collective-model-intelligence | Platform beta:

https://sinapis.ai