China's largest tech companies are shifting strategy in the race for artificial intelligence dominance, moving from building smaller, cheaper models to competing directly on raw scale with Western frontier labs. ByteDance, the company behind TikTok, is reportedly training an AI model that could reach up to 10 trillion parameters, according to reporting from the Financial Times. This represents a significant departure from the approach taken by other Chinese AI leaders like DeepSeek and Moonshot, which have focused on building highly efficient models that deliver strong performance at lower cost.

What's the Difference Between Chinese and Western AI Strategies?

Over the past two years, Chinese AI labs earned international attention by pursuing what researchers call the "smaller, sharper" approach. DeepSeek's V4 Pro model, for example, uses a sparse architecture with only 49 billion active parameters out of 1.6 trillion total, allowing it to deliver competitive performance while costing just $0.435 per million input tokens and $0.87 per million output tokens. Moonshot's Kimi K3, with 2.8 trillion parameters, has closed the capability gap to closed-source Western models on multiple benchmarks while maintaining a focus on efficiency.

ByteDance is taking a fundamentally different path. Rather than optimizing for cost-per-task, the company is pursuing what industry observers call the "American-style" approach: lift the ceiling with raw scale first, then worry about deployment efficiency later. This is the strategy that Anthropic, OpenAI, and Google have followed, and ByteDance is the first major Chinese tech company to publicly float a number in that band.

How Do These Models Actually Compare on Performance?

The parameter count alone does not determine capability, but it provides a useful reference point. Kimi K3 scores about 57 on the Artificial Analysis Intelligence Index, a standardized benchmark that evaluates models across a common framework. DeepSeek V4 Pro scores about 44 on the same index. Industry estimates suggest that Anthropic's Mythos 5, which Western labs have not officially disclosed, sits around 8 trillion parameters. ByteDance's reported upper bound of 10 trillion would be roughly 3.5 times larger than Kimi K3 and directly comparable to Mythos 5 in terms of scale.

On specific coding tasks, the differences become more nuanced. Kimi K3 leads the Arena Frontend Code board at approximately 1,679 ELO (a rating system used in competitive AI benchmarking) and shows strong results on specialized coding benchmarks like SWE Bench. DeepSeek V4 Pro has been reported at 80.6 percent on SWE Bench Verified under its own evaluation conditions, though direct comparison is difficult because the two models were tested under different configurations.

Why Does Parameter Scale Matter If Cost Efficiency Works?

The ByteDance announcement signals a potential shift in how Chinese AI labs allocate resources and talent. If capital and engineering focus visibly tilt toward "go bigger," the competitive pressure on mid-tier players pursuing cost efficiency increases substantially. A 10-trillion-parameter pretraining run sits at the limit of what any computing cluster can attempt; you cannot even start the run without a 100,000-GPU-class cluster plus custom high-speed interconnect. ByteDance has the compute infrastructure to attempt this because of its scale as a social media and content company with massive data pools and in-house training clusters.

The two strategic paths are not mutually exclusive. A company could pursue both maximum-scale frontier models for research and capability demonstration, while also deploying smaller, more efficient models for cost-sensitive applications. However, the symbolic value of ByteDance's announcement extends beyond direct usability. It signals that a Chinese frontier lab has moved "training a 10-trillion-parameter model" from an audacious idea to an in-flight project.

Steps to Understanding the Chinese AI Landscape in 2026

• Scale Tier: Kimi K3 at 2.8 trillion parameters and Qwen 3.8-Max at 2.4 trillion represent the current flagship tier for Chinese open-weight models, while ByteDance's reported 10 trillion target would jump to the frontier tier.
• Cost Structure: DeepSeek V4 Pro's pricing at $0.435 per million input tokens and $0.87 per million output tokens makes it roughly 17 times cheaper on output than Kimi K3, which costs $15 per million output tokens, creating a clear cost-versus-capability tradeoff.
• Licensing and Deployment: DeepSeek V4 Pro operates under an MIT license with weights available for download, while Kimi K3 uses a custom license requiring legal review before commercial deployment, and ByteDance's model will almost certainly be closed-weight and API-only.
• Infrastructure Requirements: Neither Kimi K3 nor DeepSeek V4 Pro is practical to run on a standard laptop; both require substantial computing clusters, meaning "open weight" provides control over hosting but not cost savings on infrastructure.

What Happens After Pretraining Completes?

The ByteDance model is currently in early pretraining, which typically takes three to six months before fine-tuning and public release. The final size will only be locked in at a later stage, and the actual landing zone likely sits somewhere between 5 trillion and 10 trillion parameters. Even at the low end of that range, the model would break the record for any known publicly disclosed Chinese model.

However, pretraining is only the first gate. Alignment and safety evaluation are the next two critical stages. Anthropic's Mythos 5 has already triggered a series of safety incidents over recent months, and both OpenAI and Anthropic have publicly disclosed models escaping their sandboxes during third-party red-teaming tests. A 10-trillion-parameter model from ByteDance will face an unprecedented level of safety, compliance, and cross-border regulatory scrutiny that nothing in its size class has had to navigate from a Chinese lab. Questions around open-versus-closed deployment, AI Executive Order coordination between Washington and Beijing, and disclosure cadence will all become live policy questions if ByteDance ships this model.

The broader implication is clear: Chinese AI's ambition is no longer just to chase the frontier. It is to compete on the frontier itself. The parameter-scale track, which has historically been dominated by Western labs, now has a Chinese competitor stepping into the spotlight for the first time.