Latest stories

De-dollarization tipping point as multipolar finance takes hold

Fujian carrier a flashy flex but not a game-changer

East looking back at its West interpreters

What one is witnessing in the production adoption of a small number of US companies has long-term implications. Models once dismissed for security or reliability issues now anchor businesses that would not switch their preferences without advantages too huge to ignore. Clearly, doubts about backdoor leakages or state subsidies fade when efficiency wins.

Chinese models are setting new norms and new transparency standards. Every major Chinese release arrives with a paper, a benchmark and weights that can be downloaded by anyone without any fees or fine-print conditions.

Architectures in motion

At the heart of this surge is one principle: do more with less. The technology underpinning this new wave is the Mixture of Experts (MoE). The simple takeaway is this: Chinese models have universally adopted MoE. They are now innovating within this framework, creating relentless new waves of efficiency.

In contrast, a few key Western models, like Anthropic's Claude, are resisting MoE entirely. Others have adopted it, but without the same obsessive focus on efficiency. In China, the race is now against each other.

The number of new directions, the speed of announcements and the quality of execution are breathtaking. The rest of the section is skippable for those not interested in the broad details of MoE.

For those interested in the details, it helps to understand the context. Most observers follow AI development in broad strokes. The first wave, in 2023, was about size. It was a race for ever-larger models.

The second wave, in 2024, was about Reasoning, as“Chain of Thought” techniques became standard. Many adopted this shift quickly. Some, like Llama and Mistral, missed a beat and suffered enduring consequences. The broad stroke of the current period is this Mixture of Experts.

This method, which originated at Google, is an efficiency strategy. The MoE turns a single massive model into a federation of smaller ones. Instead of lighting up all parameters for every query, MoE activates only the experts needed for a given task. The rest stay idle, cutting costs without shrinking total scale.

MoE's genius is that it invites variation. Once a model becomes a collection of specialists rather than a single monolith, each lab can decide how to slice and assign those specialists. Some cluster by skill, where one group of experts could be for language, another for math, and another for reasoning. Others may distribute by depth with light experts for fast answers, heavy ones for complex tasks.

Some, like MiniMax, switch experts dynamically within a single query, treating each step as a separate routing problem. DeepSeek groups its experts hierarchically, passing partial results up a tree until the best layer completes the answer.

Qwen pursues efficiency through“sparse gating,” allowing multiple tiny experts to share activation slots so that no compute ever idles. Kimi, the most recent, mixes modular routing with long-context attention, calling different experts for each stage of a reasoning chain.

These design choices sound like jargon, but they translate to visible gains. Faster models. Lower cost. Longer context windows. In one generation, Minimax has exhibited extreme sparsity, achieving top performance by activating only 10 of 230 billion parameters.

Kimi is mixing MoE with a new attention mechanism that seems to grow linearly, as against quadratically for the traditional transformers, promising extreme gains for long context. DeepSeek is re-engineering core expert-routing logic itself, while Tencent shocked everyone recently with the concept of non-fixed weights.

Chinese teams have shown that there is no single best MoE recipe-only endless ways to mix experts. Each permutation unlocks another small leap in efficiency or reasoning depth.

The wider lesson is that MoE proves how far the field still is from the theoretical limits of the transformer. The pace of progress points that current architectures are unlikely to be anywhere near a final, optimized design. The prevailing belief that scaling size alone guarantees progress is passe.

MoE is a reminder that heuristics, not laws, guide this science. There is no fixed path. Efficiency breakthroughs may come from routing, sparsity, memory, or something not yet imagined. The only certainty is that where more methods are being explored, there is a higher probability of a breakthrough. China is, right now, the center of that exploration.

Economics of openness

Open source does not mean zero revenue. It changes where revenue comes from.

Running a large model still needs compute. Hosting an open model at enterprise scale demands racks of GPUs or access to cloud capacity. The capital cost shifts from license to infrastructure. For most users, that trade-off favors paying the model's creator for API access rather than self-hosting.

That is the hidden business model of openness. Free cognition, paid convenience.

DeepSeek's 671 billion-parameter model cost under six million dollars and the same number is going around for the latest Kimi model that is going head-to-head with OpenAI's best in the popular benchmarks.

Irrespective of whether the actual training costs are what is believed or a few times more, these models are trained extremely cost-effectively. As a result, Chinese teams can charge pennies and still earn margins. Their APIs are 10 to 50 times cheaper than Western equivalents. They need fewer users to break even.

The Cost of Cognition (Per 1 Million Tokens)