# ByteDance’s 10-Trillion-Parameter AI Model vs. Mythos > Source: > Published: 2026-08-11 00:10:43+00:00 ByteDance’s Seed AI division is pre-training a model with approximately 10 trillion parameters — a scale that puts it in direct competition with Anthropic’s [Claude Mythos 5](https://www.aimagicx.com/blog/claude-mythos-5-trillion-parameter-model-developer-guide-2026), currently the largest frontier model in production. The report, first published by the [Financial Times on August 7](https://thenextweb.com/news/bytedance-10-trillion-parameter-model-mythos), has been generating significant developer debate. But the headline figure is the least interesting part of this story. ## Ten Trillion Parameters: What the Number Does and Doesn’t Mean Before the scale awe sets in, some grounding. The 10 trillion figure is a candidate upper limit — a ceiling being trained toward, not a confirmed specification. ByteDance won’t know the actual count until pre-training completes, which is expected to take three to six months. Anthropic, for its part, officially discloses zero parameter counts. The widely-cited estimate for Mythos 5 is pure industry inference — educated guessing at a model Anthropic refuses to describe by size. Architecture matters here too. Moonshot’s Kimi K3 has 2.8 trillion total parameters but only 104 billion active per token — a [Mixture-of-Experts design](https://xenospectrum.com/en/bytedance-10t-ai-model-training/) that selects a small subset of the network for each inference. A 10 trillion dense model (every parameter active, every time) and a 10 trillion MoE model are fundamentally different beasts in terms of compute cost, latency, and practical deployment. Whether ByteDance is building dense or MoE is unknown. The raw number comparison against Mythos tells you almost nothing useful. ## The Export Controls Paradox Here is where the story gets genuinely interesting. ByteDance is training a frontier-scale model while effectively cut off from the most advanced AI chips on the planet. Nvidia’s Blackwell architecture — the current generation of GPU accelerators underpinning virtually every major US lab’s training runs — is unavailable to Chinese firms under export control restrictions. The H200 situation is its own saga. The Trump administration cleared ByteDance and roughly ten other Chinese tech companies to purchase up to 75,000 H200 units each. ByteDance reportedly planned to spend around $14 billion on Nvidia hardware in 2026. That order is blocked — not by Washington, but by Beijing itself. [Chinese customs authorities told import agents the chips weren’t permitted](https://www.tomshardware.com/pc-components/gpus/tiktok-owner-bytedance-to-reportedly-purchase-usd14-billion-worth-of-nvidia-ai-gpus-in-2026), and domestic tech companies were directed to pause H200 orders. The geopolitics of AI chips now runs in both directions. So how does ByteDance train a 10-trillion-parameter model without access to leading-edge accelerators? By routing around the problem. The company has been renting compute capacity from data centers outside China — primarily in Singapore and the UAE — where US export restrictions don’t apply. Huawei’s Ascend 950PR, which entered mass production in March 2026, now runs CANN Next software with CUDA-compatible programming patterns, lowering the porting cost enough that ByteDance has placed orders. Custom ASICs co-developed with Broadcom and TSMC are reportedly in the pipeline for 2026. Export controls were designed to slow Chinese frontier AI development. ByteDance is pre-training a model that may match the world’s largest production system — on hardware the US never intended to help them build. That outcome warrants serious reexamination of the policy’s actual effect. ## The Twist: China Is Also Restricting Its Own AI There is a second layer to this story that deserves more attention than it’s getting. In July 2026, Reuters reported that China’s Ministry of Commerce held meetings with Alibaba, ByteDance, and Z.ai to discuss [restricting overseas access to China’s most advanced AI models](https://www.explainx.ai/blog/china-overseas-ai-model-restrictions-reuters-july-2026) — including models not yet released, and potentially including open-weight releases. Measures under discussion include overseas-access restrictions, criminal penalties for AI technology leaks, and limits on foreign investment in Chinese AI startups. The enforceability question on open weights is genuinely hard — once model weights are public, containment becomes nearly impossible. But the intent is clear: Beijing is treating frontier AI as a national security asset. For developers outside China, this means that even if ByteDance successfully trains a 10-trillion-parameter model, API access through Western platforms is unlikely. The competitive pressure it creates on Anthropic and OpenAI — forcing faster capability releases and better pricing — may be the primary benefit that reaches developers internationally. ## ByteDance’s Actual Advantages Strip away the geopolitics, and ByteDance has a genuinely strong position for building a capable frontier model. Its Doubao assistant is already one of China’s most-used AI products, providing both real-world deployment feedback and an immediate distribution channel. TikTok’s data at scale is a legitimate training asset. And ByteDance founder Zhang Yiming has explicitly prohibited the team from relying on AI distillation shortcuts — Seed is building original capability, not copying existing models. That discipline, combined with scale ambition, is a credible path to a competitive frontier system. ## What to Watch Pre-training wraps up somewhere in Q4 2026 or Q1 2027. Fine-tuning and safety evaluation will follow, pushing a public deployment into 2027 at the earliest. First availability will almost certainly be through Doubao for Chinese users. Western developer access looks improbable given Beijing’s current policy direction. What matters in the interim is what this story reveals: that sovereign AI infrastructure — compute, chips, models — is being built at scale outside US influence, and that chip-level export controls may be accelerating that process rather than preventing it.