Picture a CIO at a Chinese manufacturer who followed the policy signal and bought stacks of domestic AI accelerators — only to find that the open large model (大模型) they want will not load, or runs at a fraction of its rated speed. The hardware is paid for. The model is downloaded. But the two do not speak the same dialect. That awkward gap between "we bought domestic chips" and "we can actually run AI on them" is the exact problem 4Paradigm (第四范式) is now built to close.
From model vendor to compatibility layer
4Paradigm has spent a decade selling its Sage (先知) AI platform to banks, energy firms and retailers — a full enterprise AI operating system (企业级 AIOS) that turns business data into decision-making agents. What changed in 2025 is the emphasis. On 22 September 2025 the company launched ModelHub XC (信创模盒), a ModelHub plus an adaptation engine called EngineX, aimed squarely at China's信创 (domestic-substitution IT) stack.
The pain it addresses is concrete. Most popular models were tuned for Nvidia hardware; on domestic accelerators from Huawei Ascend (华为昇腾), Cambricon (寒武纪), Iluvatar CoreX (天数智芯), Kunlunxin (昆仑芯), MetaX (沐曦) or Sunrise (曦望), each model historically needed weeks of bespoke porting. EngineX pushes adaptation down to the algorithm-architecture level — Transformer, Diffusion, CNN — so that "one adaptation, many models plug-and-play." On launch day the platform listed "over a hundred" certified models, including DeepSeek V3.1 and OpenAI's open gpt-oss-20B, and 4Paradigm said it would scale certified models to the thousands within half a year and to the hundred-thousands within a year.
Why this is a company story, not just a product
4Paradigm is not a chip maker and not a foundation-model lab. Its bet is to sit in the seam between them — the unglamorous "last mile" where AI actually ships. That positioning shows up in the numbers it disclosed for the first nine months of 2025:
- Total revenue of about ¥4.4 billion (≈ US$620M), up 36.8% year on year.
- Sage platform revenue of about ¥3.7 billion (≈ US$520M), up 70.1%, now roughly 84% of group revenue.
- The company's first-ever profitable quarter, achieved in Q3 2025.
- R&D spend of about ¥1.49 billion (≈ US$210M), with the R&D-to-revenue ratio falling as revenue scaled.
These figures are company-disclosed in its HKEX filing (stock code 06682.HK) and echoed by China Securities Journal and other financial outlets. 4Paradigm also cites IDC data saying it has led China's machine-learning platform market by share for seven consecutive years — a vendor-cited claim worth reading with mild skepticism.
The "AI Agent + World Model" framing
Underneath the compute play is a clearer thesis: 4Paradigm says the future enterprise is run by a single AI brain, not a drawer of disconnected software tools. Its "AI Agent + World Model" (世界模型) path wraps an organization's processes, inventory and decisions into agents that plan and act, rather than dashboards that merely inform. To make that run on cheap, sovereign silicon, the company also released a "Virtual VRAM" expansion card it claims stretches a single card to 256GB of effective memory — a vendor spec, but one that signals where the stack is heading: let mid-range domestic GPUs punch above their weight through software.
Reading the trend
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the real moat in Chinese enterprise AI is shifting from "who has the biggest model" to "who makes domestic compute usable at scale," and adapter layers like EngineX are where the next wave of enterprise value will accrue.
Honest limitations
Every figure above for revenue, profit and Sage-platform growth is company-disclosed in a HKEX filing, not independently audited in this article's scope; treat them as management-reported. The "hundred-thousands of models in a year" target is a roadmap promise, not a delivered count — and third-party trackers already show the certified-model number moving faster than the original plan, so the trajectory matters more than any single milestone. The IDC market-share claim is cited by the company, not verified here against IDC's raw report. The Virtual VRAM "256GB / 10× RTX 4090" comparison is a vendor benchmark we did not reproduce. We have not assessed 4Paradigm's services margins, cash conversion, or how its adapter stack performs against alternatives from Huawei or startup communities.
What readers can do now
- If you run AI on domestic Chinese accelerators, test ModelHub XC's certified-model list against your own stack before committing to a porting project — the certified set is the fastest signal of what already "just works."
- Enterprise buyers evaluating Sage should ask vendors for a reference deployment in their own industry (energy, manufacturing, finance, retail) rather than a generic demo.
- Investors can track whether 4Paradigm sustains profitable quarters beyond Q3 2025; one quarter is a turning point, not a trend.
- Researchers should watch EngineX's architecture-level adapter approach as a possible template for making non-Nvidia hardware a first-class target.