Report: ByteDance Training 10-Trillion-Parameter Model, Largest in China

The Claim

ByteDance is training an AI model with up to ten trillion parameters, according to a Financial Times report citing three unnamed insiders. If accurate, this would make it three times larger than Moonshot's Kimi K3, currently regarded as China's largest AI model.

Neither ByteDance nor the FT's sources have publicly confirmed a parameter count, model name, or release timeline. Parameter counts for frontier models are rarely disclosed by their developers and are typically estimated by industry analysts or leaked through internal sources.

Scale Comparison

At ten trillion parameters, the model would place ByteDance in a similar range to Anthropic's flagship system, which industry estimates place at roughly eight trillion parameters. Anthropic has not disclosed parameter counts for any of its models, so this figure remains an unverified estimate rather than a confirmed number.

According to the FT's sources, the model is currently in pretraining, a phase that typically lasts three to six months before moving into fine-tuning and evaluation. That timeline suggests any public release, if one follows, is unlikely before late 2026 at the earliest.

Training Approach

One source told the FT that ByteDance has avoided distillation, training on outputs generated by other companies' models, for more than a year. This claim, if accurate, would distinguish the effort from allegations previously made against other Chinese AI labs accused of distilling outputs from Western models to accelerate training.

Parameter count alone does not determine model quality. Data quality, training methodology, and architecture choices all factor heavily into final performance, and none of these have been detailed in the report.

Internal Direction

ByteDance founder Zhang Yiming reportedly told the company's roughly 2,000-person Seed research team to target world-leading model capabilities over the long term. This internal directive was described by sources but has not been independently verified through official ByteDance statements.

Broader Context

ByteDance is not alone in pursuing models at this scale. xAI is reportedly training Grok variants with six and ten trillion parameters on its Colossus 2 supercomputing cluster, according to statements from Elon Musk. Neither xAI's nor ByteDance's figures have been independently confirmed through technical documentation or benchmark disclosures.

What This Means

This report describes a model still in early training, not a finished product. No benchmark scores, pricing, context window, or release date exist yet, because none have been announced. The parameter counts cited come from anonymous sources rather than company disclosures, and historical estimates of "largest model" claims have often proven difficult to verify after release.

What is notable is the trajectory: multiple labs, including ByteDance and xAI, appear to be pursuing trillion-scale parameter counts well beyond current publicly confirmed frontier models. Whether raw parameter count translates into meaningfully better real-world performance remains an open question that only becomes answerable once benchmark results and independent evaluations are available. Until ByteDance confirms details or ships a product, this remains a claim about work in progress, not a verified model release.