Kimi K3 chưa mã nguồn mở, nước ngoài đã bắt đầu xem xét lại AI Trung Quốc

TL;DR

• After Kimi K3's release, overseas discussions shifted from model capabilities to Yang Zhilin's return to China for entrepreneurship and America's talent appeal.
• Vinod Khosla pointed to US immigration policy, while Russ Salakhutdinov stated that Yang had opportunities to stay in the US but actively chose to return to China to start his company.
• Related entities: Alibaba, Tencent, Microsoft, Nvidia. Indirectly related to OpenAI, Anthropic, Hugging Face, and AI narrative assets.

After the release of Kimi K3 in mid-July, overseas tech circles quickly pivoted their discussions from parameters and leaderboards to a more easily spread question: Why did the US fail to retain Yang Zhilin?

This question has captured market attention, not just because of the founder's typical background. Kimi K3 announced an open-weight route, with officials stating that the complete model weights will be open-sourced by July 27, 2026. For enterprises, this relates to the ability to deploy models in private environments, controlling data and costs. For investors, it challenges the commercial moats of closed-source model companies like OpenAI and Anthropic.

The controversy intensified when two named parties offered different explanations. Khosla Ventures founder Vinod Khosla used Kimi K3's success to criticize US immigration policy, arguing that America is scaring away top global AI talent. Subsequently, Yang Zhilin's PhD advisor at CMU, Russ Salakhutdinov, countered on X, stating that Yang had many opportunities to remain in the US, including contact from Apple executives, but ultimately chose to return to China to start his business.

The value of Kimi K3 lies neither in proving that Chinese AI has taken the overall lead, nor in demonstrating that the US has lost the talent war. It brings a more practical issue to the surface: open models, entrepreneurial ecosystems, and talent choices are changing the pricing benchmark for global AI.

K3 makes overseas take China's open models seriously

Kimi K3 first gained attention by demonstrating visible results on high-value tasks. According to Kimi's official blog, K3 adopts a MoE (Mixture of Experts) architecture with 2.8 trillion parameters, activating 16 out of 896 experts per inference. Simply put, the model is massive, but only activates a subset of its capabilities each time to balance performance and inference costs.

Competition in large models is no longer solely about chat experience. The market cares more about whether AI can write code, break down multi-step tasks, plan long-term, and correct errors—essentially, agent tasks (autonomously completing multi-step work). These scenarios are closer to software development, enterprise automation, and future workflow gateways.

Kimi officials state that K3 supports million-token contexts and native vision capabilities, ranking highly in some coding and agent evaluations. Caveats must be noted. The company's own statements mention that K3 still lags behind some of the most advanced overseas closed-source models overall, and its performance on some leaderboards requires more independent replication.

Nevertheless, K3 has been sufficient to prompt overseas developers and investors to reassess China's open models. Previously, many Chinese models were viewed as low-cost alternatives. Now, the market is beginning to discuss whether they can enter the top tier in scenarios with higher commercial value, such as coding, agents, and long-context processing.

Open weights amplify this signal. Closed-source models are more like cloud services, accessed via APIs. Open weights allow enterprises to obtain the weights and deploy and adapt the model in local or private environments. For companies worried about data leaks, cost overruns, or vendor lock-in, this is primarily a procurement choice, not a technical stance.

Yang Zhilin's case cannot be simplified to 'America lost him'

Yang Zhilin's story easily feeds into American talent anxiety. Public information shows he earned his bachelor's degree from Tsinghua University, obtained his PhD from CMU, worked at Google Brain and Meta AI, returned to China to found Moonshot AI, and secured support from Alibaba and Tencent.

Khosla's judgment taps into a long-standing pain point in the US tech industry. H-1B visas, green cards, and geopolitical restrictions make the path to staying in the US more uncertain for international students. Top-tier labs heavily rely on global talent; the more policy friction the US creates, the more likely marginal talent flows will shift.

However, Russ Salakhutdinov's statement changed the causal chain in this specific case. As Yang's advisor, he offers a perspective close to the individual: Yang was not without opportunities in the US but believed he would regret it for life if he didn't return to China to start his company. Framing Kimi K3's success directly as talent loss due to US immigration policy lacks sufficient evidence.

A more appropriate explanation is that both forces are at play. The US immigration and international exchange environment certainly imposes more uncertainty on foreign talent. Simultaneously, domestic entrepreneurial opportunities, capital support, and the open-model approach in China are forming a strong pull. The former explains America's anxiety; the latter explains why this choice wasn't necessarily just a passive departure.

For investors, a single case cannot prove the US is systemically losing the talent war. But it sufficiently demonstrates that the optimal choice for top AI talent is no longer defaulting to staying at a major US tech company or starting a venture within the US academic system.

Open weights challenge closed-source commercial control

The greater significance of Kimi K3 is connecting the talent narrative with the model route narrative. When a returnee entrepreneur produces an open-source model with global discussion value, overseas discussions naturally extend from who trained the model to who controls the future distribution and deployment of AI.

The core advantages of closed-source model companies remain top-tier model capability, unified APIs, developer ecosystems, and enterprise customer relationships. Once open-weight models catch up in areas like coding, agents, and long contexts, they will alter the enterprise negotiation structure. Companies may not completely abandon closed-source models, but they can use open models to replace some workloads, lower costs, and retain more data sovereignty.

Tencent's Hunyuan Hy3, Alibaba's Qwen, and overseas open models are being considered along the same lines because enterprises are seeking a second infrastructure option. Tencent's Q1 earnings report mentioned that since April 28, 2026, Hy3 preview has been one of the most widely used models on OpenRouter.

These examples don't prove open weights have defeated closed source, but they do illustrate that developers' and enterprises' concerns about data control are rising. Microsoft CEO Satya Nadella has warned companies about the risks of providing data to proprietary models. Hugging Face has long advocated that open models can reduce power concentration. These aren't endorsements of a particular Chinese model but reflect the genuine concerns of enterprise customers.

For asset pricing, this affects two types of narratives. Companies like Alibaba and Tencent, which simultaneously possess models, cloud services, and application distribution, may benefit from the expansion of the open model ecosystem. Closed-source frontier companies like OpenAI and Anthropic still hold capability advantages, but their valuation narrative needs to continuously prove that their lead, enterprise stickiness, and security/compliance strengths are sufficient to offset the catch-up by open models.

Adoption rates will determine the revaluation slope

Currently, Kimi K3 provides a strong early signal, not a confirmed industry inflection point. Leading in specific benchmarks, shifting overseas public opinion, and announcing open-weight plans indicate the market is searching for a new reference system. However, these are not yet equivalent to enterprise revenue, long-term retention, or production-grade substitution.

What needs verification is whether open-weight models can transition from developer trials to enterprise deployment. Whether enterprises are willing to migrate core workflows to models like Kimi K3, Hy3, or Qwen depends on stability, inference cost, security audits, toolchain compatibility, and service support—not just leaderboard rankings.

The same applies to the talent side. Yang Zhilin has proven that returning to China for entrepreneurship can yield globally visible results. However, whether this triggers a larger-scale talent migration depends on adjustments in US policy, Chinese companies' ability to consistently provide world-class research environments, and the actual choices of the next cohort of top PhDs.

The most important aspect of Kimi K3 for investors is that it destabilizes old assumptions. Top-tier AI capabilities may not be defined solely by closed-source giants, and top talent may not necessarily need the US system to maximize their potential. Leaderboard hype fades, but whether the complete weights are open-sourced as planned, whether enterprises actually adopt them, and whether developers stay engaged—these factors will determine how far this revaluation can go.