How China's Open-Source LLMs Are Closing In on America's Closed-Source Giants

A Counterintuitive Phenomenon

By conventional logic, OpenAI, Anthropic, and Google—burning through tens or even hundreds of billions of dollars a year—should firmly hold the leading position in AI large models, backed by deep pockets, ample compute, and closed technological moats. Yet reality has taken a dramatic turn: domestic Chinese large models, represented by Kimi, Alibaba's Qwen, and DeepSeek, are not only steadily closing the performance gap but have also chosen a radically different path—fully open-sourcing their models and making them freely available to developers worldwide.

According to Bilibili tech blogger Yun Xiaozhang, this isn't a simple game of technical catch-up, but a reconstruction at the level of the rules themselves. The closed-source camp bets on "my model is the best, and you must pay to use it," while Chinese teams bet on "my model is good enough, but I'll open it up for free and win through ecosystem and scale." The two tracks are running simultaneously, and the latter is already beginning to show results.

A Hard-Power Comparison of Three Domestic Open-Source Models

Public data shows that domestic open-source models have now fully entered the trillion-parameter tier.

Kimi's flagship model boasts roughly 2.8 trillion parameters, scoring 1679 in a coding capability benchmark—surpassing Anthropic's latest version to claim the top spot—with full weights available for download.

Alibaba's Qwen new-generation model reaches 2.4 trillion parameters, the first in the series to break the trillion mark as a multimodal large model. Its code generation accuracy is said to reach 94.7%, an improvement of about 11 percentage points over the previous generation, and it's open-sourced under the permissive Apache 2.0 license.

The Apache 2.0 license explicitly grants users patent rights, permitting commercial use, modification, and redistribution—and is widely considered the most enterprise-friendly. Any company can build commercial products on Qwen's weights without worrying about patent disputes. This is especially critical in the realm of AI models: model weights involve multiple legal dimensions including copyright, patents, and data derivative rights simultaneously. Apache 2.0's explicit patent-grant clauses mean enterprises can incorporate the model into production systems without bearing the legal risk of future patent claims. By contrast, early versions of Meta's LLaMA restricted commercial use—and it is precisely this difference in openness at the license level that has profoundly influenced the pace of ecosystem expansion for each model.

DeepSeek's flagship version has roughly 1.6 trillion total parameters with 49 billion active parameters; the lightweight Flash version has 284 billion total parameters and only 13 billion active parameters, yet supports an ultra-long context window of up to 1 million tokens. A token is the basic unit by which large language models process text—roughly understood as a "word fragment," with a single Chinese character typically corresponding to 1 to 2 tokens. The context window determines the total amount of text a model can simultaneously "see" and remember in a single inference: early GPT-3 supported only about 4096 tokens, whereas 1 million tokens can accommodate over a million characters of continuous text. Achieving this capability in engineering terms requires not only expanded memory and compute but also the introduction of techniques at the attention-mechanism level such as sparse attention, sliding-window attention, and KV Cache compression—standard self-attention has computational complexity that scales with the square of sequence length, and without modification, the computation for 1 million tokens would become unbearable. This means you could feed an entire copy of The Three-Body Problem into it, and the model could remember the plot from start to finish without "crossing wires." DeepSeek is open-sourced under the MIT license, which allows free use, copying, modification, and even commercial use as long as the original license text is retained—making it one of the most widely circulated and lowest-barrier licenses in the open-source community today.

All three models are open-source across the board, all in the trillion-parameter tier—something that would have been almost unimaginable not long ago.

Why Can Domestic Open-Source Models Catch Up to Closed-Source Ones?

The MoE Mixture-of-Experts Architecture: Innovation Born of Compute Constraints

These domestic large models almost all adopt the MoE (Mixture of Experts) architecture. This architecture traces back to a 1991 academic paper, but its true revival in the large-model era came after Google released Switch Transformer in 2021. Switch Transformer scaled the number of experts to 2048, and Google's GLaM model in 2022 further validated MoE's advantages in multi-task generalization, cementing the architecture's standing in industry. Its core mechanism introduces multiple "expert" sub-networks into the Transformer's feed-forward layers, equipped with a "router" that dynamically selects which experts to activate for each token.

The router's design is the core engineering challenge in deploying MoE—early "Top-K routing" suffered from load imbalance (some experts were over-activated while others sat nearly idle). DeepSeek effectively solved this engineering pain point by introducing an auxiliary loss function to constrain routing distribution, allowing the advantages of sparse activation to be fully unleashed. This "sparse activation" property enables models to scale to enormous parameter sizes, while the actual compute during inference is tied only to the number of activated experts rather than the total parameter count—using relatively limited inference compute to leverage a parameter capacity several times larger than a Dense Model.

Taking DeepSeek as an example, only about 49 billion parameters are activated during inference.

To use an analogy: you have ten thousand employees, but for each project you dispatch only the few hundred most critical people—saving compute while preserving efficiency. Because domestic teams have long faced compute constraints, they have instead been "forced" to squeeze the maximum value out of every card, refining their MoE engineering implementation ever more skillfully. This is a textbook case of constraint breeding innovation.

The Home-Field Advantage of Chinese Data

American models are primarily built on English corpora and don't fully cover Chinese internet content. Domestic large models are naturally steeped in bilingual Chinese-English data, and their harvesting of code data is also more thorough. Qwen topping the relevant leaderboards and DeepSeek's stellar reputation in coding communities are both backed by a wealth of high-quality content produced by Chinese developers—a home-field advantage that simple translation cannot replace.

The Open-Source Flywheel Effect: The Most Devastating Competitive Weapon

The "Flywheel Effect" was systematically articulated by management scholar Jim Collins in Good to Great; at its core, it means that once initial momentum accumulates past a critical point, a virtuous cycle self-reinforces and continuously accelerates. In the open-source AI arena, this mechanism operates with particular clarity: open the weights → developers download and fine-tune → the community contributes data and feedback → model quality improves → more users adopt → commercial application scenarios emerge. Meta's LLaMA series was an early validator—after being open-sourced in 2023, fine-tuned models derived on the Hugging Face platform surpassed several thousand within months, greatly solidifying its central position in the open-source ecosystem. And the LLaMA flywheel spun so fast in no small part thanks to the permissiveness of Apache-style licensing—only when developers are confident their commercial products won't face legal risks over licensing will they truly embed a model into core business rather than keep it at the experimental stage. This collective open-sourcing by domestic models is, in essence, replicating and amplifying this playbook, layered with home-field ecosystem advantages and lower barriers to use.

When you open-source a model, millions of developers worldwide proactively download, test, and fine-tune it, feeding bugs and optimization points back to you.

Within 48 hours of Kimi's new model launch, servers were overwhelmed, forcing the company to suspend new user registrations; DeepSeek's downloads on GitHub keep setting new records. This is equivalent to the world's smartest engineers doing quality testing for free—collaboration at this scale is something the closed-source model can't buy no matter how much money it spends.

Extreme Cost Control: The Fuel That Keeps the Flywheel Spinning

The inference cost of large models is typically measured by "cost per million tokens." In early 2023, GPT-4's API pricing was roughly $30 per million input tokens and $60 per million output tokens—the industry benchmark at the time. Today, DeepSeek and Kimi's inference pricing has been driven down to extremely low levels, with cache-hit pricing as low as 0.05 RMB per million tokens (about $0.007)—a reduction of over 99%, in an entirely different league from the closed-source camp's fees.

More importantly, the low prices aren't sustained by loss-making subsidies but are the result of multiple technologies working in concert: MoE sparse activation reduces the actual compute consumed per inference; Speculative Decoding accelerates generation via a small model drafting predictions verified by the large model—this technique was systematically proposed by a Google research team in 2023, and its principle is to use a tiny "draft model" to rapidly generate several candidate token sequences, which the main model then verifies in parallel. Since verification can be executed in parallel while generation must be sequential, this "trading parallel verification for sequential generation" strategy can boost throughput by 2 to 4 times without any loss in output quality; the KV Cache sharing mechanism substantially reduces the computational redundancy of repeated requests. These techniques fundamentally lower costs rather than relying on subsidies to sustain them. This forms a virtuous cycle: lower cost → more users → more feedback data → faster model iteration → stronger competitiveness. Once the flywheel gets spinning, closed-source models will struggle to compete head-on. The collapse of the cost barrier also means that AI capabilities once affordable only to large enterprises are beginning to fully permeate long-tail application scenarios.

The Gap Remains, But Has Shifted from "Can't Catch Up" to "Catching Up Fast"

An objective gap still exists. In areas like agent capabilities, deep understanding of complex scenarios, and comprehensive information processing, America's top models still hold the lead. But the nature of the gap has changed—from the past's "can't catch up" to today's "catching up fast."

A telling signal: when the closed-source camp's strategy leaders personally step in to comment, and Anthropic is forced to lower subscription quotas, it precisely shows they're not indifferent to open source, but that they care too much and yet find it hard to stop. The open-source camp's true winning move isn't replicating the closed-source path to success, but redefining the entire rules of the game.

Conclusion: A Collision of Two Business Philosophies

The core logic of this competition has long transcended a pure technical contest. It is a head-on collision of two business philosophies: the elite route of closed and paid, versus the ecosystem route of open and free. For now, the second path has begun to bear fruit.

For developers worldwide, this may well be a good thing—stronger domestic large models, lower usage costs, and a more open ecosystem, with the ultimate beneficiary being the pace of innovation across the entire industry.

Key Takeaways