MoMA: China Mobile's LLM Aggregation Platform Putting 300+ Models Behind One Door

The short answer

MoMA is China Mobile's first open, broadly accessible large-model aggregation platform, bringing 300+ mainstream models — open-source, closed-source and the in-house Jiutian series — into one console. A unified API gateway makes a single integration enough to call them all, without registering with each vendor or adapting to each protocol. Intelligent routing matches requests to models under cost-first, quality-first or balanced strategies and fails over in seconds. RAG knowledge bases, low-code agent development and visual workflow orchestration round it out, and new accounts receive 25 million free tokens.

The kind of problem it targets

The most common blocker for enterprises adopting large models is rarely "can we use one" — it is "which one, how do we integrate it, and what will it cost". Every vendor brings its own account system, authentication method and interface protocol, so each new model means another round of adaptation work. Once live, the same task can differ substantially in price across models, and manual picking rarely balances cost against quality well.

MoMA folds that layer into platform capability: you face one gateway instead of dozens of vendors, model selection moves to a policy engine instead of case-by-case judgement, and billing is itemised per actual token consumed.

Core capabilities

Model plaza. Over 300 models spanning text, vision, speech, embedding and multimodal categories. Each supports online trial, API invocation, one-click deployment and fine-tuning, filterable by category, function or family, or reachable through direct search.

Unified API gateway. One standardised interface reaches every model on the platform. There is no need to adapt to different vendors' protocols piecemeal, nor to maintain separate accounts.

Intelligent routing (Jiutian inside). Request intent is analysed automatically and matched to the most suitable model, under cost-first, quality-first or balanced strategies. Single-point failures trigger failover within seconds, keeping production traffic continuous. The published figures are roughly 30% lower unit token cost and over 50% lower resource footprint.

RAG retrieval-augmented generation. Multi-source knowledge base construction with hybrid retrieval and multi-path recall brings scattered internal and external knowledge into the generation pipeline, reducing hallucination and improving answer accuracy.

Agent development. Two orchestration modes — LLM autonomous planning and explicit workflow — support no-code or low-code AI agents without deep programming skills.

Agent store. Ready-made agents across domains can be tried immediately, and agents you build can be published to the store as a channel for internal reuse or external distribution.

Visual workflow orchestration. Drag-and-drop composition of plugins, models, code and knowledge bases into stable processes suits multi-step business automation.

Token metering and transparent billing. Per-transaction token consumption details plus latency and throughput monitoring, with both pay-as-you-go and one-off billing options.

Getting started

First, visit the MoMA product page at ecloud.10086.cn, China Mobile Cloud's portal, register and sign in to reach the console. Second, open the model plaza to browse the 300+ models, filtering by category, function or family, or searching for a specific target. Third, pick a model and hit "online trial" to test immediately, or choose API invocation to obtain the endpoint and key. Fourth, go to the intelligent routing module, choose a strategy, tick the model set and create automatic dispatch rules so later requests follow the policy. Fifth, for enterprise question answering, upload documents in the RAG module to build a knowledge base and configure hybrid retrieval. Sixth, use the agent development module to assemble your own agent, then publish it to the agent store for teammates or external users.

How it differs from other platforms

MoMA positions itself around aggregation rather than single-vendor model service. Compared with common cloud platforms, it claims 300+ models against a typical hundred-odd; integration is built on one API covering everything while most platforms still rely on standard APIs plus SDKs; routing stresses dual-mode SLA-tiered routing with second-level failover; and infrastructure rests on China Mobile Cloud's network plus 1,500+ edge intelligence centres, paired with carrier-grade networking and content safety controls — an independent pitch on latency and compliance. If your core need is balancing cost and quality across models, an aggregator pays off directly. If you are deeply tied to one vendor's model ecosystem, that vendor's native platform will still feel smoother on supporting tooling.

Representative scenarios

Enterprise AI middle platform. A single model entry point serves multiple business lines, stopping each team from integrating vendors separately.

Knowledge Q&A and customer service. RAG turns enterprise knowledge into answers that cite their source rather than inventing content.

Agent application development. Low-code orchestration produces vertical applications such as approval assistants and data-analysis agents quickly.

AIGC content production. Combine text, image and video generation models with workflows for batch creation and publishing.

Industry-specific solutions. Fine-tune base models on industry data, then layer RAG and agents for finance, healthcare or government scenarios.

Before you commit

The platform suits engineering or IT teams that call several models and want unified billing and unified integration. If you only need one model steadily, the aggregation layer adds less. Although billing meters real tokens with itemised monitoring, cost still scales linearly with volume: use the free allowance to validate the pipeline first, then decide from routing results whether to pin to specific models. Finally, even with second-level failover, keep a fallback plan for business-critical paths.