In the summer of 2024, a researcher with a single gaming GPU could finally run a serious video-generation model at home. That was the point. Zhipu AI (智谱AI), one of China's leading large model (大模型) labs, had just open-sourced CogVideoX, and the smallest version was built to fit on hardware ordinary people already own.
The move reframed a race that had been measured in closed demos and waitlists. Instead of asking creators to line up for an API, Zhipu handed them the engine.
Why a "small" model was the headline
The first released version, CogVideoX-2B, was engineered around accessibility. In FP16 precision it needs only 18GB of VRAM for inference and 40GB for fine-tuning — which means a single NVIDIA RTX 4090 can run it, and a single A6000 can fine-tune it. For context, that is a consumer or prosumer card, not a data-center part.
The trade-offs are real but deliberate: the 2B variant produces 6-second clips at 8 frames per second, in 720×480 resolution, with a prompt limit of 226 tokens. Those numbers look modest next to cinematic showcases. But the design goal was never spectacle — it was to let a student, a small studio, or a hobbyist actually touch the technology.
Zhipu's CEO, Zhang Peng (张鹏), framed the release explicitly around lowering the barrier: a model that runs on a card someone already bought is a model someone will actually build with.
The app that proved the demand was real
Open weights are one story. Real usage is another. Zhipu bundled the model behind a consumer feature called Qingying (清影) inside its Zhipu Qingyan (智谱清言) app, launched in July 2024. The pitch was simple: type text or upload an image, wait about 30 seconds, get a 6-second video.
The uptake was immediate. According to state-media coverage from Xinhua (新华网) and Science and Technology Daily (科技日报), Qingying generated more than one million videos in its first six days of availability. That is not a benchmark score; it is evidence that text-to-video had crossed from novelty into daily habit for a large consumer base.
The family grew fast
CogVideoX was not a one-shot. Zhipu expanded the family over the following months:
- CogVideoX-5B (open-sourced late August 2024) — a larger, stronger model that the team said could run on a desktop-class RTX 3060, further pushing the "consumer GPU" thesis.
- CogVideoX-5B-I2V (open-sourced September 2024) — an image-to-video variant that takes a still image plus a prompt and animates it, adding a third task type alongside text-to-video and video extension.
- The 2B variant's license was shifted to Apache 2.0, a permissive open-source license friendly to commercial use.
By late 2024, an upgraded CogVideoX v1.5 raised output to roughly 10-second, 4K, 60-fps clips and was accompanied by companion CogSound (sound-effects) and CogMusic (music) models, extending the system from silent clips toward scored video. These later upgrades were reported by Chinese tech media and should be read as the model's continued trajectory rather than its verified launch spec.
The microdrama (短剧) connection
One detail from the original release hints at where open video models actually earn money. Zhipu noted that the 2B model, once fine-tuned, could support basic production of microdrama (短剧) content — the short, punchy serialized videos that have become a massive Chinese mobile-entertainment category.
That is the quiet commercial logic of open-sourcing: a model small enough to fine-tune on an A6000 becomes a template small studios can bend to their own house style. The infrastructure for a thousand niche video tools is exactly this kind of permissively licensed, consumer-grade checkpoint.
A view from the observatory
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the strategic value of CogVideoX was less about any single clip quality and more about "establishing a domestic open baseline that local developers stop rebuilding from scratch." That qualitative read matches the public trajectory: once a credible open model exists, the energy shifts from re-implementing to extending.
What readers can do now
- Run it locally. If you own an RTX 4090 or better, pull CogVideoX-2B from Hugging Face or ModelScope and generate clips without any subscription — a genuinely rare capability for video models.
- Fine-tune for a niche. The 40GB fine-tune footprint means a single A6000 can adapt the model to a specific visual style or product, something closed APIs rarely permit.
- Use it as a teaching artifact. Because the code, weights, and technical report are all public, CogVideoX is one of the clearest open references for how a DiT-based video model is actually assembled.
Honest limitations
This article leans on official and state-media sources — the Xinhua and Science and Technology Daily reports for the "one million videos in six days" claim, and Zhipu's own repositories for the VRAM, resolution, and frame-rate specifications. The later v1.5 figures (4K, 60fps, CogSound/CogMusic) come from Chinese tech-media reporting and were treated as directional, not launch-verified.
I have not independently rendered and scored CogVideoX against closed competitors, so no quality ranking is asserted here. Two dimensions are genuinely uncovered: the training-data sources behind the model are not detailed in the public materials, so bias and copyright questions cannot be answered from available documentation; and the real-world fine-tune results for microdrama (短剧) production are described by Zhipu as possible rather than demonstrated at scale, so I present that as a stated use case, not a measured outcome.