DeepSeek and open-source reasoning models challenge OpenAI in frontier AI performance

What happened

Open-weights artificial intelligence models have caught up to proprietary frontier systems on core reasoning and coding benchmarks, upending the long-held assumption that leading capabilities require hundreds of millions of dollars in compute.

Chinese research group DeepSeek released its R1 model family, demonstrating that pure reinforcement learning can train large language models to solve complex mathematical, logic and programming problems. The open-weights model performs on par with OpenAI's o1 across standard evaluation suites, including AIME and MATH datasets, while releasing its model weights and training methodologies publicly.

At the same time, open-source researchers and labs across the industry have begun reproducing these chain-of-thought reasoning techniques, making advanced problem-solving models accessible to independent developers and enterprise teams running local infrastructure.

Who is involved

DeepSeek, a Hangzhou-based artificial intelligence research group funded by high-frequency trading firm High-Flyer, led the recent shift by publishing its training pipeline and distilled models. The group used a mixture-of-experts architecture paired with large-scale reinforcement learning to incentivise internal reasoning tokens before generating final answers.

Major proprietary labs, including OpenAI, Google DeepMind and Anthropic, continue to develop closed reasoning architectures designed to scale compute at test-time. Meanwhile, developers in the open-source community, supported by platforms like Hugging Face, are distilling these chain-of-thought outputs into smaller models ranging from 1.5 billion to 70 billion parameters.

Why it matters

The breakthrough shifts the primary bottleneck in artificial intelligence from training compute to inference efficiency. By allowing models to think through problems step by step before answering, post-training techniques produce dramatic jumps in accuracy on hard tasks without requiring exponentially larger parameter counts.

What makes this change significant is the cost disparity. DeepSeek claims to have trained its base models for a fraction of what US frontier labs typically spend on flagship cluster runs. While exact hardware expenditure figures have not been independently audited, the availability of open weights means developers can deploy high-performing reasoning agents on their own hardware or via low-cost API providers.

This democratisation threatens the high-margin API business models of frontier labs. It also suggests that synthetic data generation and targeted reinforcement learning may matter more for future performance gains than brute-force scaling of pre-training clusters.

What happens next

Attention is turning to multimodal reasoning and agentic workflows. Researchers are applying similar reinforcement learning frameworks to vision-language models, enabling AI systems to parse technical diagrams, code repositories and user interfaces with higher precision.

Engineers are also working to reduce the latency associated with test-time reasoning. Because chain-of-thought models generate hundreds of hidden tokens before producing a response, serving them efficiently remains challenging. Compute providers are actively re-architecting inference hardware and software stacks to handle variable-length reasoning traces.

Whether proprietary labs can maintain a lead by scaling test-time compute further remains to be seen, but the baseline for open-source capability has permanently shifted.