China’s open-weight AI ecosystem is moving from developer experimentation into enterprise procurement. Alibaba, DeepSeek, and other Chinese developers now offer models with competitive capabilities, permissive licenses, and prices that challenge proprietary platforms for selected workloads.
Alibaba released Qwen3.5 in February 2026 and followed it in July with Qwen3.8-Max-Preview, a multimodal model the company says contains 2.4 trillion parameters. Alibaba also promised an open-weight release but had not provided a timeline at launch. The rapid release cycle is widening enterprise choice, but independent evidence does not support claims that China has overtaken the US across the broader AI market.
Open-weight models make trained parameters available for download, allowing organizations to run them on controlled infrastructure and adapt them for specific applications. The label does not guarantee disclosure of training data, development methods, or safety testing.
China’s open-weight push is changing the cost equation
Alibaba’s portfolio now spans downloadable Qwen3.5 releases and the newer Qwen3.8-Max-Preview. The wider Qwen family had produced more than 100,000 derivatives on Hugging Face by March 2026, making it the platform’s largest model ecosystem, according to a US-China Economic and Security Review Commission report.
Licensing lowers another barrier. Qwen3.5-397B-A17B uses the Apache 2.0 license, while DeepSeek-R1 uses the MIT License. Both permit broad commercial use, but enterprises still need to review the exact weights, distilled versions, dependencies, and usage policies included in a deployment.
Independent testing shows that capability gaps remain. A May 2026 CAISI evaluation found DeepSeek V4 was the most capable Chinese model the agency had tested, but its aggregate performance remained about eight months behind the US frontier.
DeepSeek V4 was less expensive than a comparable US reference model on five of seven benchmarks. Depending on the test, its cost ranged from 53% lower to 41% higher.
Businesses are also testing lower-priced Chinese AI models for coding, support, and internal automation. An MIT Sloan analysis of OpenRouter activity from May through September 2025 found that open-model inference averaged 23 cents per million tokens, compared with $1.86 for closed models.
OpenRouter represented about 1% of global inference spending. The comparison also excludes migration, hosting, accelerators, monitoring, security, staffing, and model-management costs.
Deployment control comes with new risks
Open weights may fit document processing, coding assistance, retrieval-augmented generation, and batch inference better than autonomous customer-facing agents or regulated decision systems. Evaluations should use the organization’s data, languages, latency requirements, and accuracy thresholds rather than general leaderboards.
Security testing should sit within a broader governance program as the enterprise AI visibility gap grows. A September 2025 CAISI assessment found agents using DeepSeek R1-0528 were 12 times more likely than tested US systems to follow malicious instructions in agent-hijacking scenarios.
In separate jailbreak testing, the model answered 94% of malicious prompts, compared with 8% for the US reference systems. Those results apply only to the tested versions and configurations.
Enterprises should test prompt injection, tool abuse, data exfiltration, harmful output, political bias, and local-language behavior. APAC organizations must also examine data residency, cross-border transfers, cloud-region availability, public-sector restrictions, and industry rules in every market where a system will operate.
Self-hosting can increase control, but telemetry, software updates, external tools, and managed services can still create outbound data flows. Adoption decisions should rest on workload performance, licensing, total operating cost, security, data movement, and jurisdictional exposure — not model origin alone.
Read more: Reported plans to restrict foreign access to advanced Chinese AI models show why enterprises need contingency plans for changes in model availability, licensing, and cross-border access.