TL;DR
ByteDance is reported to be training an AI model with 10 trillion parameters, a scale that would rank among the largest training efforts disclosed to date. The claim is attributed to the company’s ByteDance Seed research division, but ByteDance has not publicly confirmed the figure, and no architecture, timeline or benchmark details are available.
ByteDance is training an artificial intelligence model with roughly 10 trillion parameters, according to a report citing the company’s ByteDance Seed research division — a scale that, if confirmed, would rank among the largest AI training efforts disclosed by any company to date. ByteDance has not publicly confirmed the figure, and no technical documentation has accompanied the report.
According to the report, the work is being carried out by ByteDance Seed, the company’s core artificial intelligence research unit. The report did not specify the model’s architecture, its intended purpose, or a training timeline, and it remains unclear whether the figure refers to a single model or a family of models.
A parameter count of 10 trillion would exceed the largest openly documented models by a wide margin. Meta’s Llama 3.1 tops out at 405 billion parameters, and DeepSeek’s V3 model uses 671 billion total parameters in a mixture-of-experts design. OpenAI has never confirmed GPT-4’s size, though it is widely reported to use a mixture-of-experts architecture of roughly 1.8 trillion parameters in total.
The raw figure alone says little about capability. Modern frontier models often use mixture-of-experts designs, in which only a fraction of parameters handles any given input. A 10-trillion-parameter model built this way could activate far fewer parameters per query, which would keep computing costs more manageable in everyday use.
ByteDance Is Reportedly Training a 10-Trillion-Parameter AI Model
A report citing ByteDance’s Seed research division claims the TikTok parent is training a model at a scale that would rank among the largest AI efforts ever disclosed. The figure is unconfirmed — no architecture, timeline, or benchmarks have been released.
Status: Reported claim · Source: ByteDance Seed · Confirmation: None
How 10 Trillion Stacks Up
If confirmed, the figure would exceed the largest openly documented models by a wide margin. Note: GPT-4’s size has never been confirmed by OpenAI, and total parameters do not equal active compute per query.
Frontier Models at a Glance
The reported ByteDance effort against the best-documented large models — with verification status colour-coded.
| Model / Lab | Total Parameters | Architecture | Disclosure Status |
|---|---|---|---|
| ByteDance Seed model | ~10T | Unknown — dense or MoE unspecified | ~ Reported only |
| GPT-4 / OpenAI | ~1.8T | Widely reported mixture-of-experts | ✗ Never confirmed |
| DeepSeek-V3 | 671B | Mixture-of-experts | ✓ Openly documented |
| Llama 3.1 / Meta | 405B | Dense | ✓ Openly documented |
What a 10T Run Would Signal
A training run at this scale would mark a sharp break from the efficiency-first strategy that has defined Chinese AI labs — a “scale first” bet with commercial and geopolitical weight.
Competing at the very top
Few laboratories anywhere — even in the U.S. — have approached this scale. It would signal ByteDance intends to compete at the frontier, not just ship cheap models for consumer apps.
Fueling Doubao and beyond
Doubao is among China’s most-used AI products. A stronger foundation model could feed advertising, video generation, and every app in ByteDance’s ecosystem.
Pressure on DeepSeek & Alibaba
Alibaba’s Qwen models and DeepSeek’s efficiency breakthroughs have set the pace for Chinese AI. A 10T effort would sharpen that competition considerably.
Tens of thousands of chips
Runs of this size typically require tens of thousands of advanced chips — while U.S. export controls restrict top-end Nvidia hardware for Chinese firms.
MoE could tame inference
If built as mixture-of-experts, only a fraction of parameters would activate per query, keeping everyday computing costs far more manageable.
Among China’s top AI buyers
Analysts tracking regional chip procurement rank ByteDance among China’s most aggressive spenders on AI infrastructure.
From ChatGPT Shock to Scale-First
ByteDance’s AI build-out, traced from the Seed division’s founding to the reported 10-trillion-parameter run.
⚡ Early 2023 · Seed founded
Core AI research unit created months after ChatGPT’s debut jolted China’s tech industry.
🤖 Late 2023 · Doubao launches
Flagship chatbot grows into one of China’s most widely used consumer AI products.
What a 10-Trillion-Parameter Model Would Mean
If the report is accurate, ByteDance would be attempting a training run at a scale few laboratories have approached, even in the United States. That would signal that the TikTok parent company intends to compete at the frontier of AI development rather than focus only on smaller, cheaper models for its consumer apps.
The stakes are both commercial and geopolitical. ByteDance’s Doubao chatbot is among the most used AI products in China, and a stronger foundation model could feed everything from advertising to video generation across the company’s apps. The effort would also sharpen competition with domestic rivals such as DeepSeek and Alibaba, whose Qwen models have set much of the pace for Chinese AI over the past year.
Training at this scale raises cost and supply questions as well. Runs of this size typically require tens of thousands of advanced chips, and Chinese companies face U.S. export controls on top-end Nvidia hardware. How ByteDance would source or substitute that computing power is one of the biggest practical questions hanging over the report.
ByteDance’s AI Push Behind Doubao and Seed
ByteDance established its Seed division in early 2023 to work on large language models and related research, months after ChatGPT’s debut jolted China’s technology industry. The company launched Doubao later that year, and the chatbot has since grown into one of China’s most widely used consumer AI products.
The Seed team has since released a series of models under the Doubao brand, spanning text, image and video generation, and has published some of its research openly. ByteDance has also been among China’s most aggressive spenders on AI infrastructure, according to analysts tracking chip procurement in the region.
Chinese labs have so far competed largely on efficiency — DeepSeek drew global attention by training capable models at a fraction of typical costs. A 10-trillion-parameter effort would mark a different strategy: scale first.
“ByteDance is training an AI model with 10 trillion parameters.”
— The report, citing ByteDance Seed
Key Details the Report Does Not Confirm
Almost everything beyond the headline number is unconfirmed. The report does not specify whether the model is dense or mixture-of-experts, what data it is being trained on, how far training has progressed, or what the model is intended to do. It is also not clear whether the figure refers to one model or several. ByteDance has not verified the claim, and no technical paper, benchmark result or official announcement has accompanied the report. Readers should treat the 10-trillion figure as a reported claim until the company confirms it.
Signals to Watch From ByteDance’s Seed Team
The clearest signal would be an official statement or research release from ByteDance Seed. The team has a track record of publishing papers and releasing some work openly, so technical documentation — if it comes — would likely appear through research channels first. New Doubao-branded model releases, benchmark submissions, and hiring or chip-procurement disclosures are the other markers to watch in the coming months. Until then, the 10-trillion figure remains a reported number rather than a confirmed specification.
Source: ByteDance Seed
Key Questions
What did the report about ByteDance’s new AI model actually say?
The report says ByteDance is training an AI model with 10 trillion parameters, attributing the work to the company’s Seed research division. It provided no confirmed details on architecture, timeline or purpose, and ByteDance has not publicly verified the figure.
Has ByteDance confirmed the 10-trillion-parameter model?
No. The figure comes from a report citing ByteDance Seed. The company has not issued a public confirmation, technical paper or benchmark data to support it.
How does 10 trillion parameters compare with other major AI models?
It would be far larger than openly documented models. Meta’s Llama 3.1 reaches 405 billion parameters, DeepSeek-V3 uses 671 billion, and GPT-4 is reported — though never confirmed by OpenAI — at roughly 1.8 trillion in a mixture-of-experts setup. Total parameter count also does not equal active compute per query, which matters for cost.
What is ByteDance Seed?
Seed is ByteDance’s core AI research division, established in 2023. It develops the foundation models behind Doubao, the company’s flagship chatbot, as well as image and video generation systems.
When could this model be released?
No timeline has been reported. Watch for research papers, benchmark submissions or new Doubao-branded releases from ByteDance Seed as the first public signs of the project.
Source: ByteDance Seed
AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
GIGABYTE AORUS RTX 5090 AI Box Graphics Card – External GPU (32GB GDDR7, 512-bit, PCIe 5.0, HDMI/DP 2.1b, 240mm Radiator, Silent Fans, Direct-Coverage Copper Plate, Thunderbolt 5™)
- High-Performance GPU: Powered by GeForce RTX 5090 with NVIDIA Blackwell architecture
- Advanced Cooling System: Waterforce all-in-one with copper base and radiator
- Quiet Operation: Two silent 120mm fans for thermals
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Local AI Engineering with Ollama: Run, understand, customize, fine-tune, and build agentic apps on your own hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
