Alibaba Unveils Qwen 4, Zhenwu V900 and AgentCore Roadmap
The Apsara Conference plan connects a model now in training with a 2027 accelerator target and enterprise agent services, while most performance claims remain vendor-reported.
Edited by Tyronne Panaino
Alibaba announced a full-stack AI roadmap at its Apsara Conference in Hangzhou on September 22, spanning models, proprietary chips, cloud infrastructure and agent software. The company said Qwen 4 is in training, unveiled the Zhenwu V900 accelerator and described AgentCore and Agent Context as parts of an enterprise agent stack.
The update matters to model builders, cloud buyers and device makers because Alibaba is presenting these components as one connected system rather than isolated releases. Several pieces are available or newly unveiled, while other parts remain future plans with limited independent evidence.
Qwen 4 is a training update, not a model release
Alibaba said Qwen 4 is currently in training. It also outlined Qwen 4.5 and Qwen 5 series that the company projects could scale to between 5 trillion and 10 trillion parameters. Those later model names and sizes are roadmap statements, not shipping products or independently verified specifications.
The announcement also introduced Qwen Intelligence, a business-facing agent platform for smartphones. Alibaba says the platform is intended to give phone manufacturers a Qwen-based foundation for cross-application tasks. The source does not name launch partners, supported devices or a general availability schedule, so the immediate significance is the product direction rather than proven consumer reach.
Zhenwu V900 targets a 2027 commercial release
T-Head, Alibaba's chip unit, unveiled the Zhenwu V900 for AI training and inference. Alibaba lists 216 GB of memory and 1,200 GB per second of inter-chip bandwidth, with support for multiple numerical formats including FP8 and FP4. The company schedules mass production and commercial release for the first quarter of 2027.
That timing separates the V900 from software announced for nearer use. Alibaba also described a supernode design around the processor and a CPU roadmap for agentic workloads, but the retained source provides no customer pricing, independent performance testing or firm deployment commitments. The specifications should therefore be read as vendor claims ahead of commercial hardware.
AgentCore and context services form the enterprise layer
Alibaba divides its agentic-cloud plan into model, harness and context layers. AgentCore is described as a platform for building, running and managing AI agents throughout their lifecycle, while an Agent Security Center is intended to add monitoring, security and compliance controls.
Agent Context is the data layer in that design. Alibaba says it connects documents, business systems, conversations and multimodal information so agents can use real-time context and long-term memory. The announcement also places training and inference services in an AI Native Cloud layer, making the practical delta the breadth of the proposed stack: model development, accelerator hardware, orchestration, security and context management are being presented under one cloud strategy.
What the evidence establishes
The official announcement establishes that Alibaba made these product and roadmap disclosures on September 22. It does not independently validate the company's benchmark improvements, performance multipliers, cost reductions or maximum cluster-scale claims. It also leaves important availability details unresolved for Qwen 4, Qwen Intelligence and several cloud services.
The next verifiable checkpoints are a Qwen 4 release with technical documentation, named availability terms for the agent services and commercial V900 hardware in the first quarter of 2027. Independent evaluations and customer deployments would be needed to judge whether the integrated stack performs as described.
Status
Confirmed company announcement; medium internal confidence. The event and roadmap are documented by Alibaba, while performance, scale and cost claims remain vendor-reported.
Sources
Update note: Last reviewed 2026-09-22. We will revise this post when Alibaba publishes model documentation, service availability details or commercial V900 evidence.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.