ODPS 2026 Upgrade: AI-Native Multimodal Big Data Platform for Agentic Era

ODPS Strategic Upgrade for Agentic AI Era

At the 2026 Yunqi Conference "Agent-Driven Multimodal Big Data Computing Innovation" forum, Alibaba Cloud announced a comprehensive strategic upgrade of its self-developed big data platform ODPS, reconstructing data infrastructure for the Agentic AI era. The three core products — ODPS-MaxCompute, ODPS-Hologres, and ODPS-DataWorks — were jointly released with agent-native architecture and multimodal computing capabilities at their core, providing next-generation data infrastructure for scenarios such as large model data refinement, embodied intelligence, autonomous driving data processing, and enterprise intelligent operations.

Two Fundamental Shifts in Big Data Computing

In the opening dialogue, Datafun founder Wang Dachuan, Alibaba Cloud VP Wang Junhua, and AMD China GM Liu Hongbing discussed the co-evolution of AI compute and big data computing. Wang Junhua summarized two profound changes: the object of computation has changed — expanding from structured tables and logs to images, audio/video, sensor data, and other multimodal data; and the subject of computation has changed — more tasks are initiated, orchestrated, and executed by Agents, so the platform serves not only humans but also AI. Liu Hongbing added that customers no longer treat big data platforms and AI platforms as separate "compute islands," demanding generality, heterogeneous collaboration, and open extensibility. Both agreed that big data and AI are merging into a single integrated system — the central thesis of this ODPS upgrade.

01 ODPS-MaxCompute: Everything Is Computable

MaxCompute lead Zhang Zhiguo presented the core releases. As multimodal becomes mainstream for GenAI, the data processing unit shifts from a single record to a video clip or audio-visual segment, and traditional ETL cannot handle the complexity of unstructured data. Around the philosophy "everything is computable," MaxCompute underwent full-stack reconstruction across storage, compute, engine, and interaction.

SQL AI: Closed-Loop Vectorization and Retrieval Inside the Warehouse

The most visible change is that a single SQL statement can do much more. In multimodal data processing, a common pain point is moving data across multiple systems: first embedding for vectorization, then building indexes and retrieving in external systems — long chain, high cost, complex O&M. MaxCompute SQL AI newly adds EMBEDDING and VECTOR_SEARCH capabilities, closing the loop from vectorization to retrieval entirely inside the warehouse , with zero data egress and zero resident services. This release marks another key milestone in MaxCompute SQL AI evolution: from 2023 MaxCompute 4.0 proposing Data+AI, to 2024 Object Table bringing unstructured data into SQL, to 2025 AI Function multimodal operators enabling direct model invocation, to 2026 heterogeneous compute support, vector retrieval closed-loop, and full multimodal processing — SQL AI has traversed the complete path from "connecting to AI" to "native AI" in four years.

Concrete results: Zhiyi Technology migrated tens of millions of product image vector matching from a standalone retrieval cluster to MaxCompute SQL AI, achieving 73% CPU reduction, 93% memory reduction, and 5.86x retrieval acceleration . Qwen APP in RAG corpus ingestion evaluation saw full-link evaluation time for 24 million data records and 1.2 million queries shrink from 6 hours 15 minutes to 32 minutes, with 80% cost reduction .

Multimodal Storage and Heterogeneous Compute Foundation

Underlying these scenarios is MaxCompute's new multimodal storage and heterogeneous compute system. Storage layer: Blob large object type uses reference-entity separation ; text data auto-compresses saving ~5x space on average; Object Table manages unstructured data on OSS in tabular form with cache acceleration and query optimization ; together they enable single-row multi-column mixed storage and unified management of text, image, audio/video, 3D, sensor, and other multimodal data . Compute layer: CU Quota (general CPU) supports AutoScaling intelligent scaling; GU Quota provides serverless GPU elastic on-demand use, single-job thousand-GPU concurrent scheduling, and automatic faulty GPU replacement; Inference Quota connects to Alibaba Cloud Bailian commercial LLMs with per-token metered billing . Three heterogeneous resource types switch on demand, realizing true pay-as-you-go Data+AI fused computing. MaxCompute also deeply integrates Bailian commercial LLMs; SQL + MaxFrame dual engines unify access, and through TPM/RPM rate-limiting auto-concurrency splitting and backoff strategies, large-scale inference capability is pushed to the limit under the same rate limits.

MaxFrame Upgraded to Distributed AI Compute Engine

MaxFrame is fully upgraded as a distributed AI compute engine; DPE heterogeneous resource unified scheduling supports 100k CU cluster-scale capacity, with CPU/GPU/Token three resource types mixed scheduling and automatic per-operator matching . Qianche Intelligence collected fisheye video and 6DoF pose data via UMI handheld gripper, then performed pose reconstruction, semantic annotation, and LeRobot dataset generation. Using MaxFrame AI Function and serverless heterogeneous compute, data processing throughput increased 10x+, elastic compute peak extended to 100k+ CU, model training efficiency improved ~50%, with zero manual intervention throughout . Yiwu Commodity City used MaxFrame AI Function to understand hundreds of millions of product images and videos; million-CU elastic scaling doubled performance versus self-built solution, while integrating Qwen and DeepSeek models for one-click multilingual product description globalization. Notably, MaxFrame simultaneously released industry Skills, encapsulating senior engineers' development experience into reusable AI Skills, drastically lowering the barrier to building industry data pipelines.

Agentic Interaction Suite

On the interaction layer, MaxCompute launched a full-scenario Agentic suite: AI Query natural language querying, AI Coding assisted coding, MaxAgent full-link intelligent O&M . MaxAgent executes as the user identity without modifying data, covering six O&M scenarios — job diagnosis, cost optimization, permission audit, etc. — replacing 30 minutes of manual troubleshooting with 30 seconds. The simultaneously released MCP Server wraps platform APIs as Agent-callable tools, enabling third-party Agent plug-and-play. In the Data Agent direction, MaxCompute built a platform-hosted semantic layer, precipitating metadata, lineage, business terminology, and metric definitions into SemanticSpec business specifications; Agents fetch context on demand, reducing repeated trial-and-error; underlying O&M semantic packages let Agents go straight from problem to action.

02 ODPS-Hologres 5.0: From Real-Time Data to Reliable Action

Hologres lead Jiang Weihua noted that OLAP users are shifting from humans to Agents. Previously, business personnel asked ad-hoc SQL/BI questions, manually interpreted, then manually drove next steps; now Agents run continuously, autonomously combine context, and feed results back into the next round of judgment. Agents bring lower barriers, continuous operation, and closed-loop action, but also pose four new challenges to OLAP: real-time complete data, understandable semantics, trustworthy controllable actions, and adaptive execution against unpredictable queries. Hologres 5.0 upgrades to an Agent-Native multimodal real-time data warehouse around "from real-time data to reliable action," addressing each challenge.

Data Sandbox: Let Agents Experiment Safely

Jiang believes the biggest obstacle for Agents entering enterprise data pipelines is not insufficient intelligence but trust. Previously Agents could only read data, not write, because a wrong write impacts production. Hologres 5.0's newly released Data Sandbox creates an isolated writable workspace based on the production instance; Agents can freely execute CRUD and DDL changes for experimentation, and after verification and approval, merge back to the source instance without affecting production . The live demo showed this capability: faced with "omni-channel promotion GMV decline," the Agent completed a four-step closed loop in the sandbox — locate low-conversion channels, experiment with metric definition changes, replay historical data to compare old vs. new definitions, generate pending-approval jobs for merge-back — turning a multi-round manual investigation into a single system closed loop.

Graph Query and Semantic View: Let Agents Understand Business Meaning

Another trust dimension is enabling Agents to truly understand business data semantics. Hologres 5.0's new graph query capability moves the warehouse from "find similar" to "query along relationship paths"; based on AGE extension, it calls Cypher subset via SQL and combines with regular SQL, so relationship chains like "customer → places order → order → contains → product" can be expressed directly in SQL, supporting knowledge graph construction and Agent decision-making . The simultaneously released Semantic View builds a business semantic layer inside the warehouse; after declaring table relationships and metric definitions, the engine automatically rewrites multi-table JOINs and guarantees aggregation semantic correctness, letting humans, BI, and AI use the same business language, making LLM data retrieval more accurate .

User Defined Agent and Hologres Agent: Embedding AI Logic in Data Pipelines

On Agent definition and execution, Hologres 5.0 introduces User Defined Agent, allowing users to define AI processing logic in Python and embed it into data pipelines for continuous automatic execution . Users define the Agent, then trigger AI inference row-by-row via DataFrame, and use Dynamic Table incremental materialization for auto-refresh — the engine processes only new rows as source tables receive new data, no full recomputation needed. The same Agent logic supports both batch and streaming processing. Example: product review AI scoring — each new review yields an AI score and rationale in the scoring table within 1 minute. Hologres also releases Hologres Agent — a built-in console intelligent assistant that uses LLM capabilities to complete product consultation, instance O&M, data warehouse development, and data analysis via natural language . On this basis, Hologres 5.0 officially launches three intelligent agents: O&M Agent, Tuning Agent, and Data Agent, targeting system O&M, performance tuning, and data development scenarios respectively, sharing real-time data, Skills, and security boundaries.

Three-Pillar Trust Foundation: Rememberable, Experimentable, Observable

Around the three values "rememberable, experimentable, observable," Hologres 5.0 builds a complete trust foundation for Agents: Data Sandbox makes Agents "experimentable" — the isolation experiment and approval merge mechanism lets Agents touch data without touching production; Hologres Long-Term Memory Service makes Agents "rememberable" — achieving 96.82 on LOCOMO Benchmark (SOTA), saving >90% tokens; HoloScope makes Agents "observable" — supports 100+ metric ingestion and full-session replay evaluation . All three are built on the Hologres base and serve all general Agents. On performance, Hologres continues to set industry records: VectorDBBench vector retrieval performance leads the industry; TPC-H 3TB world first , providing solid performance guarantees for unpredictable Agent-era query patterns.

03 ODPS-DataWorks: Bringing Agents into Enterprise Data Production Lines

DataWorks product lead Tian Qixi pointed out that enterprise data processing is shifting from "people using tools" to "Agents running autonomously," but getting Agents into production lines is harder than imagined: same field name different meanings, same meaning different names; Agent-generated SQL looks correct but picks the wrong table — "Agents are smart but lack business common sense." Addressing this, DataWorks releases the enterprise semantic knowledge graph Context Graph, which does not require manual ontology and model building upfront; instead it auto-scans and infers from lakehouse/warehouse data assets, covering six semantic layers: metadata enhancement, metrics and terminology, relationships and lineage, scheduling and quality, intelligent learning, and organizational knowledge . It updates incrementally as assets change and is shared by all Agents. Before integrating the semantic graph, a cross-system data analysis required serial confirmation across multiple systems and repeated context switching; after integration, a single question automatically brings in context, achieving 93.24% business Q&A accuracy, 99% SQL generation consistency, lower model cost, with Qwen3.8-flash performance approaching flagship model levels .

DataWorks accumulates over a decade of data development practice: daily scheduling tasks 40M+, managed data scale EB-level, 15+ official expert suites, 80+ skills covering data integration, development, O&M, governance, analysis end-to-end, cloud-hosted 7×24 online, supporting DingTalk, Feishu, WeCom multi-channel access. The newly released ADA (AI-Native Big Data Service) further enables multi-agent collaborative orchestration and cross-engine task autonomous delivery, covering offline, real-time, streaming, and AI computing full scenarios .

"Trustworthy" is the final threshold for enterprise Agent adoption. DataWorks builds three defense lines: data security line guards identity, fine-grained permissions, security sandbox, and sensitive data protection; behavior audit line controls Agent actions with full-link observability, Agent data lineage and watermark traceability; cost control line prevents credit runaway . It also offers multiple Agent engine choices, opens six extension capabilities including MCP, OpenAPI, 200+ Skills, and launches proactive intelligent O&M inspection service, shifting from human passive patrol to Agent autonomous discovery, root-cause analysis, repair, and daily report output. DataWorks Data Agent was rated Leader in the 2026 IDC China Data Agent Vendor Assessment.

From Cloud Infrastructure to AI-Native Multimodal Big Data Infrastructure

Over sixteen years, ODPS has evolved from supporting Alibaba internal businesses to becoming universal cloud infrastructure, and now to comprehensive reconstruction for the Agentic AI era — each evolution tightly tracking the pulse of technology development. Alibaba Cloud, through its Agent-native, multimodal computing intelligent data platform, is continuously lowering the threshold for AI technology adoption, empowering thousands of industries to unlock data value in the Agentic AI era, moving from real-time data to reliable action.