The week of October 26–November 2, 2026 marked a concentrated expansion of China’s AI stack: Baidu released a 280-billion-parameter model, Meituan launched a 1.6-trillion-parameter multimodal system, chip portability moved higher on product roadmaps, and regulators widened their attention from models to agents, employment, exports, and accelerators.
Here is our professional analysis of this week’s most critical developments in Chinese AI. For CTOs and engineering leaders, the signal is clear: model selection in China now requires simultaneous evaluation of capability, hardware compatibility, regulatory exposure, and commercialization readiness.
Baidu: A 280 Billion-Parameter Efficiency Play
ERNIE 5.5 Activates 14 Billion Parameters per Token
Baidu released ERNIE 5.5 Flash Base, which Recode China AI described as a 280-billion-parameter mixture-of-experts base model with 14 billion active parameters per token. The same report said the model supports context lengths of up to one million tokens during training and inference.
That architecture places the operational emphasis on sparse activation rather than total parameter count alone. The headline is 280 billion, but the infrastructure decision turns on how the 14-billion-parameter active path affects memory, latency, and serving costs under your actual workloads.
Why this matters: Enterprise teams should not compare Chinese models through total parameter counts alone. Your evaluation plan needs separate measurements for active parameters, context utilization, inference throughput, and deployment hardware. A million-token ceiling is not automatically a million-token production requirement.
This is not merely a scale contest — it is an efficiency contest conducted inside a constrained compute environment.
Meituan: A 1.6 Trillion-Parameter Multimodal Model
LongCat-2.5 Targets Long, Multistep Workflows
Meituan launched LongCat-2.5, which Recode China AI described as a natively multimodal model with 1.6 trillion total parameters. The model is designed for long, multistep tasks, including coding workflows that incorporate visual feedback.
China AI Bulletin reported that Meituan trained LongCat-2.5 entirely on domestic hardware. That disclosure is strategically important because it connects frontier-model development with a local compute supply chain rather than treating hardware independence as a future objective.
For engineering leaders, multimodal coding changes the evaluation surface. A system that interprets visual feedback alongside code must be tested for state tracking, action sequencing, interface recognition, and recovery from incorrect intermediate steps—not just code completion quality.
Why this matters: Teams considering LongCat-2.5 should design workflow-level evaluations rather than prompt-level demonstrations. The practical question is whether the model can complete a controlled sequence reliably across text, code, and visual interfaces. This is agent infrastructure, not a larger chatbot.
MiniMax Extends Support Across Nine Chip Platforms
Meanwhile, MiniMax released M3.1 Flash in versions adapted for nine chip platforms, according to China AI Bulletin. This expands the competitive boundary from model quality into deployment portability.
Supporting nine platforms can reduce dependence on a single accelerator family, but each adaptation creates a validation obligation. Your team still needs to measure numerical consistency, throughput, memory behavior, and operational tooling on the exact platform selected for production.
Perhaps the most consequential shift is that domestic hardware support is becoming part of the model release itself. DeepSeek is also reportedly planning to pretrain its next model on Huawei Ascend chips, according to Recode China AI. If completed, that would deepen the alignment between Chinese frontier-model development and domestic accelerators.
Moonshot, Ant, and Tencent: Open Weights Meet Agents
Kimi K3.1 Pro Adds Cyber-Data Filtering
Moonshot AI published the weights for Kimi K3.1 Pro after filtering cyber-related training data, according to Recode China AI and China AI Bulletin. The release expands access to the model while making dataset governance part of the publication decision.
Tencent also released Hy4 with visual reasoning and computer-use capabilities. China AI Bulletin reported that Tencent characterized its cyber training data as defensive and said internal security laboratories audited the model.
These are vendor-reported controls, not independent security findings. Enterprise adopters should treat them as inputs to due diligence rather than substitutes for red-team testing, permission boundaries, and monitored execution environments.
This is not conventional model governance. It is agent governance, where outputs can become actions.
Ling-3.2-Flash Scales to 560 Billion Parameters
Ant Group introduced Ling-3.2-Flash, a 560-billion-parameter multimodal model aimed at visual reasoning, 3D interaction, and medical applications, according to Recode China AI.
Those target domains carry different failure costs. A visual reasoning benchmark cannot establish suitability for medical use, while performance in 3D interaction does not prove safe operation in an enterprise application.
What this means for IT leaders: Separate evaluation tracks are required for each intended domain. Access controls, human review, logging, and escalation policies should follow the risk of the workflow—not the model’s broad multimodal label.
Regulation: Oversight Moves Across the AI Lifecycle
Employment and Agent Reviews Enter Project Planning
China’s National Development and Reform Commission said it would accelerate comprehensive AI legislation and develop lifecycle safety standards and ethical rules, according to Recode China AI. China AI Bulletin also reported that the State Council’s General Office is requiring authorities reviewing some AI projects to consider employment risks, particularly where major job displacement may occur.
The Central Cyberspace Affairs Commission called for stronger security assessment and filing requirements covering AI agents, on-device models, and applications in key sectors. Draft export-control changes would also require registration before certain listed AI technologies are exported, according to China AI Bulletin.
For CTOs, compliance planning can no longer begin at model launch. It must cover training data, model capabilities, agent permissions, deployment sector, workforce effects, and cross-border distribution.
New market-regulation standards and guidance also address how AI accelerators connect and exchange data. That brings the underlying compute layer into the same policy perimeter as models and applications.
Why this matters: Architecture reviews should record model provenance, accelerator dependencies, agent permissions, intended users, and export pathways. These are becoming governance fields, not optional documentation.
Commercialization: MaaS Revenue Rises 9.6 Times
Revenue Reaches RMB 10.6 Billion
China’s model-as-a-service revenue reached RMB 10.6 billion in the first half of 2026, up from RMB 1.1 billion in the same period of 2025, according to IDC data cited in a Futu industry preview. That represents approximately 9.6 times the prior-year level.
This is not merely expanding model availability — it is evidence of a commercial delivery layer forming around APIs, managed inference, and enterprise consumption.
Apple also reportedly received approval to introduce Apple Intelligence in China through partnerships involving Alibaba’s Qwen and Baidu, according to InjazAI. The reported structure reinforces the importance of local models and compliance partners for foreign technology companies operating in the market.
Alibaba, meanwhile, reportedly previewed Qwen3.8 Max as a 2.4-trillion-parameter model, with broader open-weight access planned, according to Stratechery. Planned access is not a release, so procurement teams should wait for published weights, licensing terms, and deployment requirements before treating it as an available option.
The direction is nevertheless consistent: larger models, broader modality, domestic compute, and tighter governance are advancing together. China’s AI market is building a parallel operating stack—models, chips, services, and rules—rather than competing on model benchmarks alone.
This weekly analysis is based on recent market developments, technical releases, and policy announcements shaping the trajectory of AI in China.