The race for larger parameters in large language models is back on, with a new funding round for Beijing-based AI startup Moonshot AI (DeepSeek) quickly spreading through the industry on July 29. The company secured over $3.5 billion in its latest financing, pushing its post-investment valuation to $35 billion, and has already initiated a Pre-IPO round with a pre-investment valuation of $50 billion. This marks a more than tenfold increase in valuation from its $4.3 billion valuation during the Series C round at the end of 2025, a pace rarely seen among global AI startups. A request for comment on the funding details was sent to Moonshot AI but no response was received by the deadline.

The core catalyst for this valuation surge was the release of Moonshot AI's flagship model, Kimi K3, on July 16. The model features approximately 2.8 trillion parameters (2.78 trillion to be precise), making it the largest open-source model globally by parameter count. On July 19, Moonshot AI announced a temporary halt to new individual user subscriptions, channeling all its computing power to serve existing users. On the evening of July 27, the company further released the model weights for Kimi K3 along with a 47-page technical report detailing the model architecture and training methods.

The launch of Kimi K3 is just one example of a broader shift in China's large language model landscape toward larger parameters. Other major models include Alibaba's flagship Qwen3.8 Max, released on July 19, with about 2.4 trillion parameters; Baidu's Wenxin 5.0, launched in January, also with 2.4 trillion parameters; and DeepSeek's V4-Pro, released in April, with 1.6 trillion parameters. Additionally, insiders report that MiniMax (00100.HK) will have over 2 trillion parameters in its next-generation model. In 2025, the release of DeepSeek R1 led the industry to briefly pivot toward smaller-parameter, lower-cost approaches, sparking a debate over whether "Scaling Law" was becoming obsolete. However, just a year later, the "parameter race" has reignited, with China's leading model developers collectively aiming for models with over 2 trillion parameters.

Parameter Leap

The scale of parameter increase for Kimi K3 is uncommon in the industry. Compared to its predecessor, Kimi K3's total parameters surged from 1.04 trillion to 2.78 trillion, a 167% increase. Its activated parameters soared from 32.6 billion to 104.2 billion, a 220% jump, and its context window expanded from 128K (K for thousand) to 1 million tokens, nearly an eightfold increase. In the technical report for Kimi K3 released on July 27, the Moonshot AI team noted that over the past year, open-source models have made rapid progress in reinforcement learning during the inference stage, but the parameter scale of pre-training foundations has stagnated around 1 trillion.

Training a large language model involves two stages: pre-training, where the model learns language, knowledge, and reasoning patterns from massive data to build foundational capabilities, with parameter scale and data quality determining the upper limit of these capabilities; and post-training, where the development team uses methods like reinforcement learning (using reward signals to guide the model in improving its reasoning strategies) and chain-of-thought reasoning (breaking down the reasoning process into multiple steps) to refine the capabilities built on top of the foundation. Over the past year, the industry's primary focus has been on post-training, as DeepSeek R1 demonstrated that reinforcement learning can achieve good results on models with smaller parameters, leading both domestic and international companies to follow suit. However, the Kimi K3 technical report argues that the optimization space for post-training is constrained by the capacity of the pre-training foundation, with the capability boundary of the foundation determining the upper limit of post-training.

As early as March at the Zhongguancun Forum, Moonshot AI founder Yang Zhilin publicly articulated a similar view. He argued that building large models is essentially about converting more energy into intelligence, and that scaling is not merely about increasing computational power but is a competition for conversion efficiency. In his speech, he outlined three dimensions for improving efficiency: enhancing token efficiency to extract more intelligence from limited data; optimizing long-context architectures to handle complex tasks over extended periods with low loss; and introducing an Agent cluster model (multiple AI agents working in parallel) to expand the boundaries of complex tasks the model can execute.

A semiconductor analyst at a major Chinese brokerage firm told the Economic Observer that the resumption of the parameter race also has an external driver: starting in late June 2026, leading overseas models like Claude began restricting domestic models' access to their APIs. This means that distillation (using the output of a large model to train a smaller one), a key path for domestic models to narrow the gap over the past two years, is becoming increasingly difficult. In his view, the development of large language models can be divided into three phases: Phase 1, from early 2023 to the third quarter of 2024, focused on pre-training parameter competition; Phase 2, starting in the second half of 2024, shifted to post-training reinforcement learning and inference optimization; and Phase 3, starting in 2026, involves model capabilities evolving from single-agent intelligence to system-level orchestration, with AI transitioning toward Agent forms. He believes Phase 3 does not replace the earlier phases, as the continuous growth of parameter count remains a key variable. China's pursuit of leading models will primarily rely on three paths: aligning training paradigms with overseas models, building self-sustaining data feedback loops, and continuously expanding the parameter scale of pre-training foundations.

A product manager at an internet gaming company in Shenzhen told the Economic Observer that for users, the direct impact of parameter scale is the model's ability to handle complex tasks. He noted that his team has integrated AI programming tools into their daily development workflow since early this year. In a complete programming task, the model needs to read the entire project context, continuously call tools for compilation, testing, and debugging, and perform multiple rounds of self-correction. The token consumption for a single session in such scenarios is many times higher than in traditional dialogue scenarios. This type of long-range task places demands on the model's underlying capabilities that are on a completely different level from simple question-and-answer interactions. Consequently, "how long a model can work continuously" is becoming a new metric for measuring model capability.

Scaling Law

Despite its 2.8 trillion parameters, Kimi K3 does not need to use all of them for every inference. According to the technical report, Kimi K3 employs a Mixture of Experts (MoE) architecture. The core idea of this architecture is to divide the model's parameters into multiple functional modules called "experts." When processing each input, a routing system activates only the small subset of experts most relevant to the current task, while the rest are not involved in the computation. Kimi K3 has 896 expert modules, activating only 16 during each inference, resulting in a sparsity factor of 56. The actual number of activated parameters involved in computation is about 104 billion. In other words, a 2.8 trillion-parameter model, with a single inference, has a computational cost roughly equivalent to a dense model (a traditional model that uses all parameters for every operation) at the hundred-billion-parameter level. The MoE architecture addresses the problem of "large parameter count but manageable inference cost."

Building on this, Moonshot AI also needed to solve two other challenges: the computational overhead in long-context scenarios and the layer-by-layer decay of information in deep networks. According to the Kimi K3 technical report, Moonshot AI developed two foundational architectures to address these. The first is KDA (Kimi Delta Attention, a hybrid linear attention mechanism). When processing long contexts, traditional Transformer architectures need to maintain a cache whose size grows linearly with the input length. A context of 1 million tokens means enormous memory usage and computational cost. KDA compresses this ever-expanding cache into a fixed-size matrix state, significantly reducing the computational cost for million-level contexts. The second is Attention Residuals. In traditional deep networks, information passes from lower layers upward, losing some information at each layer. Attention Residuals allow each layer to selectively look back at the outputs of all previous layers, bypassing the limitations of sequential transmission. According to the Kimi K3 technical report, the additional cost of this improvement is about 2%, but it significantly mitigates the problem of information decay in deep networks.

With the combination of MoE, KDA, and Attention Residuals, Kimi K3's overall expansion efficiency improved by about 2.5 times compared to Kimi K2, meaning that with the same computational investment, Kimi K3 achieves better model performance. This efficiency gain is also reflected in usage costs. According to public API pricing, Kimi K3's input price is 20 yuan per million tokens, and the output price is 100 yuan. In US dollars, Kimi K3's output unit price is about $14 per million tokens, lower than GPT-5.6 Sol's $30 and Claude Fable 5's $50. On third-party evaluations, Kimi K3's average single-task cost is about $0.94, roughly equivalent to GPT-5.6 Sol and about half of Claude Opus 4.8. In the pursuit of "producing more intelligence with less computation," edge-side model companies seem to be moving faster. Li Dahai, CEO of ModelBest AI, mentioned a set of comparative data in an interview: his company's edge-side model, with 1 billion to 2 billion parameters, has achieved the level of GPT-4o (OpenAI's flagship multimodal model released in 2024, with parameters in the hundreds of billions) from two years ago, representing nearly a hundredfold compression. ModelBest AI co-founder and President Lei Shengtao calls this direction the "knowledge density law," where the capability density of edge-side models doubles approximately every 3.5 months.

Cloud models are expanding parameter scale, while edge-side models are compressing it. The underlying logic is the same: to increase the intelligence output per unit of computation. In Li Dahai's view, scenarios like cloud-based programming and long-range Agent applications are beginning to take off, while edge-side applications haven't yet seen a similar commercial explosion, but model capability improvements are ongoing. Two years ago, achieving GPT-4o's level required models with hundreds of billions of parameters, but now ModelBest AI can achieve it with edge-side models ranging from 1 billion to 2 billion parameters. Cloud models are getting bigger, and edge-side models are getting smaller, but the fundamental law driving both is the same: more computing power and better architecture lead to stronger model capabilities. "The absence of an explosion doesn't mean there is no Scaling Law," he said. A research report from Nomura Securities on July 27 noted that 3 trillion parameters are seen by the industry as the next milestone for Chinese large language models, but model competitiveness will increasingly depend on architectural efficiency, post-training optimization, and long-range task performance.

Industry Chain Transmission

As the parameter race resumes, the commercialization of leading model developers is also accelerating. Public information shows that Moonshot AI's ARR (Annual Recurring Revenue) exceeded $300 million in mid-June, with API revenue accounting for over 70%. In a public event in June, Moonshot AI's enterprise business head Huang Zhenxin mentioned that since the release of Kimi K2.5 in January, the number of overseas paid users for Kimi has grown fourfold, and the product has entered over 200 countries and regions. Industries like internet, finance, manufacturing, education, and healthcare have become important sources of enterprise clients for Kimi. The surge in demand following Kimi K3's release has also highlighted the shortage of computing power. Public information indicates that deploying Kimi K3 requires an 8-card B300 server, and training requires a cluster of tens of thousands of cards. Several industry insiders interviewed mentioned that most of the computing power in smart computing centers has been locked up by leading tech giants, leaving independent model companies with relatively lagging access to computing power, making it difficult to replenish resources even in the face of demand explosions like Kimi K3's.

Li Dahai provided an observation from the edge-side perspective, summarizing the division of labor between edge and cloud in eight characters: "Edge manages internal, Cloud manages external." The edge holds users' private data and real-time needs, with data staying on the device, handling privacy-sensitive local inference; the cloud accesses external information and services, handling open-ended tasks that require higher model capability. He believes that as cloud model parameter sizes continue to grow, the coexistence of edge and cloud will be a long-term pattern. The semiconductor analyst mentioned earlier noted that the domestic large language model competition has formed two tiers: Alibaba and DeepSeek lead in model capability and business coverage, followed closely by Tencent, ByteDance, Zhipu AI, and Moonshot AI. The focus of competition is also shifting from underlying model capability to the comprehensive capability of the Agent OS (Agent Operating System, the system layer above the model responsible for task orchestration, tool calling, and long-range execution).

While the domestic competition landscape is rapidly changing, overseas markets are becoming another source of growth for China's leading model developers. In the aforementioned research report, Nomura mentioned that DeepSeek's ARR as of June was about $520 million, with overseas markets contributing about 47% to 48%. However, its users are primarily individual developers and small teams, with limited penetration among large enterprises. Furthermore, in real development workflows, many overseas developers have formed a habit of layered calling: using overseas models like Claude Code and Codex for architectural design and complex reasoning, then switching to domestic models like DeepSeek for ongoing code generation and execution. In other words, domestic models have initially captured the most token-intensive execution phase. A researcher at a major public fund in East China noted that the overall cost of tokens is on a downward trend, with an expected annual decline of 30% to 50%. Notably, moderately complex tasks in the Agent chain consume over twenty times more tokens per call than traditional dialogue scenarios, while ultra-complex tasks can consume millions of tokens per call. Falling costs are stimulating demand for more complex tasks, with overall profits driven by the incremental market.

The impact of "larger parameters" on the computing power industry chain is also beginning to be felt at the chip level. LinkStart Technology, an edge-side AI inference chip company incubated by the iFLYTEK ecosystem, has shipped over 150 million chips cumulatively and just completed a nearly 500 million yuan Series B funding round in July. Its first edge-side large model inference chip, the Nebula series, is planned for tape-out by the end of the year. LinkStart Technology's Market Vice President, Han Chaoyang, told the Economic Observer that the scale of the edge-side inference market will eventually be more than an order of magnitude larger than the cloud market. He also emphasized that computing power utilization is more critical than absolute peak computing power: a chip nominally rated at 200 TOPS (200 trillion operations per second) but with a utilization rate of only over 10% will have a vastly different real-world output compared to a chip rated at 100 TOPS with a 70% utilization rate. However, current edge-side chips still cannot meet the demands in terms of inference performance and power consumption, indicating that the market is still in its early stages. LinkStart Technology Vice President Xu Yan Song mentioned that chip companies and algorithm companies may no longer exist independently in the future, and chip design must invest in R&D ahead of algorithm iteration. "The chip development cycle is typically two to three years, so you must anticipate the direction of model architecture two to three years in advance, or the chip will be outdated by the time it is released," he said. This means the impact of the resumed parameter race will not remain at the model level; it is already being transmitted along the upstream industry chain, including chip design and computing power infrastructure.

In the view of the public fund researcher, the penetration of large models into industry scenarios can be divided into three phases: Phase 1 is AI programming agents, targeting the global market of about 30 million software developers and some outsourced procurement, with a core scale of about $700 billion; Phase 2 involves process-oriented AI agents, serving about 370 million white-collar office workers in areas like customer service, finance, and administration, with a total compensation pool exceeding $1 trillion; Phase 3 and beyond will extend to specialized fields like finance, healthcare, and law, with a market size of over $3 trillion to $4 trillion. He believes the industry is currently only in the early stages of Phase 1. According to this framework, the current explosion in token consumption around AI programming scenarios is still far from true large-scale penetration.