Under fireworks and dragon parades, a quieter battle is unfolding: the race to build the next generation of artificial intelligence.
While crowds mark Lunar New Year in China with lion dances and drones in the sky, the country’s tech giants are using the holiday as a launchpad for a wave of new AI models. These systems are no longer just catching up with Silicon Valley – on some fronts they are starting to press uncomfortably close.
From festive robots to a sharper ai rivalry
This year’s New Year celebrations came with a futuristic twist. Humanoid robots performed tightly coordinated choreographies alongside human dancers, a made‑for‑TV symbol of Beijing’s ambition to turn AI into a core industrial engine.
Behind the spectacle, Chinese firms quietly timed major AI announcements to the holiday window. ByteDance, Alibaba, Zhipu AI, DeepSeek and Moonshot AI all pushed out new models, many of them aimed directly at rivals like OpenAI’s GPT series and Google’s Gemini.
The US embargo on advanced AI chips was meant to slow Beijing down. Instead, it is forcing China to build leaner, more efficient models.
OpenAI chief executive Sam Altman has already called China’s progress “remarkable” in a recent interview with US television. His tone was cautious rather than celebratory: Washington’s attempt to choke off Nvidia‑class chips is colliding with a reality in which Chinese labs are learning to do more with less.
The quiet advantage: open source and local control
One of the biggest strategic differences sits where few consumers look: the licence.
Many headline Chinese models are released as either open source or “open‑weight”. That gives developers and companies the option to download the models and run them entirely on their own hardware, with no data sent back to a central provider.
Open source vs open‑weight, in practice
- Open source: full model code is published, including architecture and often training scripts.
- Open‑weight: the underlying code and training data stay closed, but the trained “weights” are downloadable.
- Result: both can run locally; only the degree of transparency differs.
This matters in two ways. First, cost: once the hardware is in place, usage is essentially free beyond electricity. Second, privacy: sensitive documents, industrial know‑how and personal information never need to cross a border or touch a US‑based platform.
For European manufacturers, banks or health providers, a capable Chinese model that lives on‑premise can be more attractive than a slightly better US system running in a distant data centre.
➡️ This plant that can stop mold is becoming the ideal natural solution for bathrooms and damp rooms
➡️ This country keeps breaking green records while Europe falls behind
➡️ Sophie Adenot unveils her space playlist: here are the artists accompanying her in orbit
➡️ This tiny habit helps you stay consistent without motivation
➡️ Meet Mississippi Mud Potatoes: The One-Pan Dish That Rivals a Loaded Baked Potato
➡️ A shock is coming: farmland values set to plunge 60% in these regions over the coming decades
➡️ In Canada, a wolf stuns scientists by outsmarting a human fishing device to get food
This dynamic is starting to worry Silicon Valley. Open models erode the high‑margin API business that underpins the current AI boom. And China’s decision to lean into openness – even if partly tactical – is giving it soft power among developers worldwide.
Seedance 2.0: cinematic video, legal headaches
The most eye‑catching launch around Lunar New Year came from ByteDance, owner of TikTok. Its new video generator, Seedance 2.0, pushed social media into a brief frenzy with clips that looked as if they had been lifted from big‑budget films.
Seedance 2.0 stands out for two reasons. The model is closed, unlike many of its Chinese peers. And its realism has already triggered serious copyright concerns. Studios including Disney, Paramount and Netflix have questioned whether their content was used in training without permission.
The dispute mirrors the scrutiny facing OpenAI’s Sora and other generative video tools. But it also sends a clear message: China is not just matching the West in text chatbots, it is competing head‑on in high‑end visual media where Hollywood’s interests are most exposed.
Qwen3.5, glm‑5 and kimi: a new wave of chinese models
Alibaba’s qwen3.5 goes multimodal
Alibaba’s research arm announced Qwen3.5, a vision‑language model designed to operate across text, images and video in around 200 languages. It can power chatbots, but also act as a multimodal “agent” able to read PDFs, fill in web forms or navigate corporate intranets.
Qwen3.5 is released under a free licence, downloadable from open repositories. Engineers can fine‑tune it for specific domains – legal contracts, logistics documents, customer support workflows – then deploy it inside their own infrastructure.
GLM‑5: efficiency built on huawei chips
Zhipu AI, one of China’s most ambitious AI labs, unveiled GLM‑5. The model is pitched for “agentic intelligence” and complex, multi‑step reasoning – exactly the territory US firms are targeting for software automation and coding assistants.
Technically, GLM‑5 rests on DeepSeek Sparse Attention (DSA), a mechanism that limits which parts of a sequence the model attends to at each step. The goal is to cut computation while preserving or even improving performance.
GLM‑5 was reportedly trained entirely on Huawei Ascend chips, sidestepping US‑made semiconductors and signalling that China’s AI stack can, at least in theory, run on domestic hardware alone.
The choice is as much political as technical. If Chinese models can hit frontier performance without Nvidia GPUs, Washington’s leverage over the country’s AI trajectory looks weaker.
Deepseek v4 and kimi k2.5: aiming at gpt and claude
Perhaps the most closely watched release has yet to fully land. DeepSeek’s previous model, V3, made waves last year by matching or approaching ChatGPT‑grade benchmarks while costing far less to train. V4, expected within days, is rumoured to be particularly strong in programming tasks.
Industry newsletter The Information reported that DeepSeek V4 outperforms Anthropic’s Claude and some GPT variants on internal benchmarks, especially for code and formal reasoning. These claims still need broad independent testing, but they show how tight the race has become.
| Model | Key focus | Openness |
|---|---|---|
| Qwen3.5 (Alibaba) | Multimodal, 200+ languages, web‑agent tasks | Open source |
| GLM‑5 (Zhipu AI) | Multi‑step reasoning, agentic use cases | Open source |
| DeepSeek V4 | Coding, logical reasoning, efficiency | Details pending, earlier versions open‑weight |
| Kimi K2.5 (Moonshot AI) | Mixture‑of‑experts, general chatbot | Partially open |
| Seedance 2.0 (ByteDance) | High‑quality video generation | Closed |
Moonshot AI’s Kimi K2.5, launched in late January, uses a “mixture of experts” (MoE) architecture similar to what Google employs in Gemini 3.0 Pro. Instead of one monolithic neural network trying to handle everything, MoE splits the model into specialised subnetworks and activates only a subset for each request.
This design cuts compute needs and makes it easier to scale up total parameter counts without exploding inference costs – a direct answer to the GPU supply crunch.
Chip bans and the efficiency arms race
While US players pour billions into vast data centres packed with top‑tier Nvidia hardware, their Chinese rivals have to think differently. Export controls mean fewer cutting‑edge chips and more reliance on second‑best or domestic alternatives.
That constraint is pushing Chinese labs toward aggressive optimisation: sparse attention, MoE routing, quantisation and highly tuned inference stacks. Over time, such techniques can spread globally, lowering the barrier for smaller countries and companies to run frontier‑like models on modest clusters.
If the US builds the fastest race cars, China may end up designing the most efficient engines – and those engines can be copied everywhere.
The risk for Silicon Valley is that dominance in raw performance morphs into a disadvantage in price‑performance. In many commercial settings, “good enough” plus low cost beats “best in class” with a hefty cloud bill.
A shrinking gap with chatgpt and gemini
On standard benchmarks, OpenAI’s GPT‑4‑class systems and Google’s top Gemini models still hold a narrow lead across languages and reasoning tasks. Yet the margin is shrinking, especially once cost, latency and deployment flexibility are factored in.
Chinese models can now run fully offline on high‑end workstations or compact server racks. That appeals to industries which care less about absolute peak performance and more about sovereignty, auditability and long‑term control.
From a European or Asian corporate perspective, the choice is no longer “use an American platform or lag behind”. It is increasingly “which trade‑off between performance, jurisdiction and openness fits our risk profile?”
What all this means for businesses outside china
For CIOs in London, Berlin or São Paulo, the rise of Chinese models adds both opportunity and complexity.
- More bargaining power with US providers on pricing and licence terms.
- New options to build in‑house tools without shipping sensitive data abroad.
- Fresh legal and geopolitical questions about adopting software linked to Beijing.
Some Western regulators are wary of embedding Chinese AI deep into critical infrastructure. Yet many models can be forked, audited and modified once open‑sourced, blurring the line between “foreign” and “domestic”. Over time, codebases may become so mixed that national origin is almost impossible to define cleanly.
Key concepts readers keep hearing about
What “agentic” ai really refers to
Several Chinese labs market their models as built for “agentic intelligence”. In practice, this means systems that do more than answer questions. They can break a goal into steps, call tools like databases or browsers, write and run code, then adapt based on feedback.
For example, an agentic AI might receive a task such as “analyse our January sales drop and propose three budget‑neutral fixes”. It could pull spreadsheets, run simple models, generate slides and draft emails – not perfectly, but well enough to save hours of manual work.
Why open‑weight models matter for ordinary users
Open‑weight releases sound niche, but they have concrete consequences. A mid‑sized law firm, for instance, could download GLM‑5, fine‑tune it on redacted case files and host it on an internal server. Lawyers get instant drafting help without sending confidential material to a US or Chinese cloud provider.
Schools and universities can run smaller variants on campus hardware, giving students powerful assistants without handing over behavioural data. Local governments might deploy translation and form‑filling bots tuned to regional dialects and procedures, all under their direct control.
Those kinds of grounded, low‑glamour deployments rarely trend on social media. Yet they are exactly where China’s bet on open, efficient models could reshape the balance of power with Silicon Valley over the next few years.
Originally posted 2026-02-05 21:21:41.