For years, the AI industry operated on a single assumption: build a bigger model. The scaling laws appeared straightforward. More parameters, more data, more compute during training — and performance would follow. That era has largely ended. Pre-training has plateaued. The internet's high-quality text has been mostly consumed. The marginal return on each additional billion parameters is diminishing.
But the AI industry did not stop scaling. It just changed when and how it scales. The new frontier is not training. The new frontier is inference.
Test-time compute scaling—giving a model more time and tokens to “think” during generation—has become the defining battleground of AI in 2026. China has built a comprehensive playbook for winning it.
The Shift from Pre-Training to Test-Time
The logic of test-time scaling is straightforward: instead of spending hundreds of millions of dollars to train a larger model, give a smaller model more compute at inference time. Let it deliberate, plan, and check its own logic before returning an answer. Research has shown that a smaller model with a much larger inference compute budget can outperform larger models running standard inference.
At the 2026 World Artificial Intelligence Conference, Wuqian Xiandong co-founder and CEO Xia Lixue articulated the shift clearly: AI is moving through four scaling laws — pre-training, post-training, test-time scaling, and agentic scaling. The industry has entered the test-time era.
Chinese AI labs did not merely adopt this paradigm. They have systematically extended it.
The DeepSeek Blueprint
DeepSeek was among the first to demonstrate test-time compute scaling at scale. R1-Zero's training revealed that the model's “thinking process” naturally lengthened during training, with AIME pass@1 rising from 15% to 79% as a result. The model exposes its full reasoning chain in `<think>` tags, making the thinking process transparent.
DeepSeek's V4 API combines a thinking toggle with graduated `reasoning_effort` levels — low, high, or max. This is not a black box. It is a tunable dial. The company has signaled R2 is anticipated for 2026.
The cost performance is striking. V4 Flash with Best-of-5 self-verification reaches 88% on Terminal-Bench 2.1 — frontier-model accuracy at roughly 1/11th the cost. The method is simple: generate multiple candidates from a cheap model, then use the same model to grade pairs of candidates. The cheap model generates great answers; it just needs a mechanism to recognize which one is great.
Beyond DeepSeek: The Ecosystem
DeepSeek is not alone. A dense ecosystem of Chinese research labs, universities, and tech companies has built an extensive test-time scaling infrastructure.
Ant Group presented EGSS at ACL 2026, addressing the fundamental inefficiency of existing test-time scaling methods. Traditional TTS applies uniform compute to every reasoning step, wasting tokens on deterministic operations while under-exploring critical decision points. EGSS uses “tool entropy” to identify high-uncertainty decision points, allocating compute only where it matters. The result: on SWE-Bench-Verified, GLM-4.6 + EGSS reached 74.6%, a new open-source record, with K=4 beating K=8 while saving 38-42% of tokens.
OPPO AI Agent team published a systematic framework for scaling test-time compute for LLM agents, laying out four key strategies: parallel sampling, sequential revision, search, and verification. This is the first systematic application of TTS to agent architectures.
StepFun, Tsinghua University, and Peking University collaborated on PaCoRe, a parallel collaborative reasoning framework that decouples test-time compute scaling from context window limitations. An 8B model using PaCoRe achieved 94.5% on the HMMT 2025 benchmark, surpassing GPT-5's 93.2%. The framework achieves effective TTC of approximately 2 million tokens within standard context limits.
ByteDance and Peking University proposed In-Place TTT, a plug-and-play test-time training method that improves long-text performance without changing model architecture, accepted as an ICLR 2026 Oral paper.
Fudan University, Peking University, and Meituan's LongCat team proposed TDAR, a Block Diffusion inference model test-time scaling framework.
Qwen3-Max-Thinking introduced a test-time scaling mechanism with dynamic compute allocation, allowing the model to conduct multiple rounds of self-verification and path search before outputting a final answer.
GenPRM demonstrated that a 1.5B process reward model outperforms GPT-4o through test-time scaling, and a 7B model surpasses Qwen2.5-Math-PRM-72B.
SCALE achieved accuracy improvements of up to 13.75 percentage points on AIME25 through selective resource allocation during test-time scaling.
The pattern is consistent: Chinese labs are not just using test-time scaling. They are systematically optimizing it, making it more efficient, more targeted, and more cost-effective.
The Chinese Advantage
The Chinese approach to test-time scaling differs from the U.S. approach in three key ways.
First, efficiency over brute force. U.S. labs often apply uniform test-time compute across all tasks. Chinese labs are developing targeted methods—EGSS, PaCoRe, TDAR—that allocate compute only where it matters, achieving the same or better results with far fewer tokens.
Second, open infrastructure. The models that power these methods are overwhelmingly open-weight. Kimi K3, Qwen3.8-Max, DeepSeek V4, GLM-5.3-Flash—all are available for developers to inspect, modify, and deploy. This is not a closed ecosystem. It is a shared infrastructure that accelerates the entire field.
Third, cost architecture from the ground up. Chinese models are built for efficiency from the architecture level. Mixture-of-Experts architectures, attention mechanism optimizations, and low-precision training all reduce per-query costs. The result is a test-time scaling ecosystem where the marginal cost of additional thinking is not prohibitive.
What This Means
The test-time scaling playbook that Chinese AI labs have assembled is not a single breakthrough. It is a systematic infrastructure—a dense network of research, engineering, and open-source collaboration that has turned test-time scaling from an academic curiosity into a production-grade capability.
The U.S. has OpenAI's o1 and Google's Gemini. China has DeepSeek R1, Qwen3-Max-Thinking, Kimi K3, and an ecosystem of researchers and engineers who are systematically making test-time scaling cheaper, faster, and more efficient.
The question is no longer whether test-time scaling works. It does. The question is who can deploy it at scale, at low cost, across the widest range of tasks.
On that measure, China's lead is increasingly difficult to ignore.
Sources: Ant Group EGSS, ACL 2026; OPPO AI Agent “Scaling Test-time Compute for LLM Agents”; StepFun/Tsinghua/Peking University PaCoRe, arXiv:2601.05593; DeepSeek AGI Roadmap Issue #1365; GitHub llm-as-a-Verifier-dsh; ByteDance/Peking University In-Place TTT, ICLR 2026 Oral; Qwen3-Max-Thinking; GenPRM, AAAI 2026; SCALE, AAAI 2026; 2026 World Artificial Intelligence Conference.
Disclaimer
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations
This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.
Forecasts from third-party analysts can change with market conditions.
Cost and pricing examples are point-in-time estimates; actual rates vary.
Country and company comparisons rely on public reporting, not operational data.
This sector moves fast; timelines and deal terms may be updated later.
Company deals and regulatory rulings may evolve; verify current status.
AI infrastructure is changing quickly; claims can become outdated soon.
Sources
- Ant Group EGSS, ACL 2026
- OPPO AI Agent “Scaling Test-time Compute for LLM Agents”
- StepFun/Tsinghua/Peking University PaCoRe, arXiv:2601.05593
- DeepSeek AGI Roadmap Issue #1365
- GitHub llm-as-a-Verifier-dsh
- ByteDance/Peking University In-Place TTT, ICLR 2026 Oral
- Qwen3-Max-Thinking
- GenPRM, AAAI 2026
- SCALE, AAAI 2026
- 2026 World Artificial Intelligence Conference.
The information provided in this article is for general informational and educational purposes only. It does not constitute legal, financial, or professional advice. The author and publisher are not responsible for any actions taken based on the content of this article. Readers should consult qualified professionals for advice specific to their situation. All trademarks and references to third-party products, services, or organizations are the property of their respective owners. The performance data and benchmarks discussed are based on specific research studies and may not generalize to all use cases or environments. As of the publication date, the AI landscape continues to evolve rapidly, and readers should verify current information independently.
Limitations: This analysis is based on reporting and public data available as of the article date; figures may be revised as sources update.; Forecasts from third-party analysts can change with market conditions.; Cost and pricing examples are point-in-time estimates; actual rates vary.; Country and company comparisons rely on public reporting, not operational data.; This sector moves fast; timelines and deal terms may be updated later.; Company deals and regulatory rulings may evolve; verify current status.; AI infrastructure is changing quickly; claims can become outdated soon.