Most Chinese labs ship video generation as a separate product: a model for images, another for video, a third for voice. Baidu took a different path with Ernie 5.0 (文心5.0), arguing that the interesting frontier is not a better video clip but a single model that understands and produces text, image, audio, and video in one unified space. Whether that bet pays off depends less on any one demo and more on whether "omnimodal" survives contact with real production.
Ernie 5.0's formal version went live on 22 January 2026 after a November 2025 preview, and Baidu has since framed it as the backbone for a wave of multimodal products — including a real-time interactive digital human built to mimic a well-known Chinese tech commentator.
Key takeaways
- Ernie 5.0 (文心5.0), released 22 January 2026, uses a native unified omnimodal architecture and 2.4 trillion parameters.
- It handles text, image, audio, and video as inputs and outputs within one model, rather than bolting modalities together after training.
- Baidu states the model reached 30 frames per second of video generation on a cluster of consumer 4090 GPUs, and that it surpassed Gemini-2.5-Pro and GPT-5-High on 40+ benchmarks.
- The same architecture powers a real-time interactive digital human (the "罗永浩 digital human"), using a three-state token-linked design for low-latency expression and lip-sync.
One model, not a toolbox
The technical claim is architectural. Where many systems train separate text and vision models and fuse them late ("late fusion"), Ernie 5.0 maps text, image, audio, and video into one semantic space under a single autoregressive framework, trained jointly. Baidu argues this removes the "semantic gap" that appears when modalities are stitched together after the fact, and lets the model reason across types — for example, taking a product photo and producing both a description and marketing copy, or turning a tutorial video into runnable front-end code.
For video specifically, that means generation is not a side quest but a native output the base model can produce. Baidu states the model can generate video at 30 frames per second on a 4090 GPU cluster, a figure worth treating as a company claim rather than a benchmarked standard, but one that signals where Baidu wants the capability to sit: inside the general model, not in a separate studio tool.
The digital human as the proof point
The most concrete deployment is a real-time interactive digital human. Baidu describes a "three-state token-linked" architecture that streams text, voice, and video together, driving expression and lip-sync from voice tokens with low latency and low cost. A digital human modeled on Luo Yonghao (罗永浩), a prominent Chinese tech commentator, was built on this technology — a recognizable test of whether an omnimodal base model can sustain a convincing, interactive persona rather than a pre-rendered clip.
This is where the omnimodal thesis meets a real market. China's live-commerce and customer-service sectors consume enormous volumes of presenters; a low-latency, low-cost interactive digital human is a direct cost lever. Baidu's bet is that owning the unified base model makes these personas cheaper and more flexible than chaining separate speech, vision, and animation models.
Where it sits in the China race
Ernie 5.0 is not alone in chasing unification — Alibaba's Qwen family and others are pushing multimodal roadmaps too — but Baidu's emphasis on a single autoregressive architecture is a distinct wager. On the purely generative-video leaderboards that creators watch (where Alibaba's Wan and Kuaishou's Kling trade the top spots), Baidu has positioned Ernie more as an infrastructure layer than a headline clip-maker. That may help it in enterprise adoption, where unified APIs beat boutique quality, and may hurt it in consumer mindshare, where a striking video travels further than a architecture diagram.
According to Li An, chief scientist at BrainNet (脑机网), the more interesting question for buyers is not which lab's demo looks best, but which unified model actually holds consistency across a long, multi-shot task without a human in the loop.
Honest limitations
This article draws on Baidu's own official materials (Baidu Smart Cloud, Baidu's corporate newsroom, and Baidu Baike), treated as the primary — and in places only — sources. Company-stated figures (2.4 trillion parameters, 30 fps on 4090 clusters, surpassing Gemini-2.5-Pro and GPT-5-High on 40+ benchmarks) are reported as Baidu states them and were not independently re-measured. The "late fusion versus unified architecture" comparison reflects Baidu's framing. I have not run Ernie 5.0, have not verified the 30 fps claim under controlled conditions, and have not compared its video output against Wan, Kling, or Seedance. The digital-human deployment is described from Baidu's account, not from hands-on testing.
What readers can do now
- If you evaluate foundation models, separate "supports video" from "does video well." A unified model that lists video as an output is different from a dedicated video engine tuned for cinematic quality.
- For interactive presenters, probe latency and cost, not just looks. The real differentiator for digital-human use is real-time responsiveness and price per session.
- Watch the architecture debate play out in procurement. Buyers should track whether unified models actually reduce integration cost versus best-of-breed single-modality tools before consolidating.