Chinese artificial intelligence has concentrated three major movements in just a few days. Moonshot AI had to restrict new subscriptions to Kimi K3 due to capacity limitations, DeepSeek has begun the general rollout of its V4 family, and Alibaba has announced Qwen 3.8 Max. All three models compete with US proposals such as GPT-5.6 and Claude Opus 5, but the real battleground is no longer just benchmarks: it also matters how much it costs to serve each response and how many GPUs are available to meet demand.
The key points of the new Chinese models in 20 seconds
- Moonshot temporarily halted new sign-ups for Kimi K3 after saturating part of its inference capacity.
- DeepSeek V4 combines long context, reduced prices, and two variants for different workloads.
- Alibaba introduced Qwen 3.8 Max as a multimodal model with 2.4 trillion parameters.
- GPT-5.6 and Claude Opus 5 maintain advantages in several professional tasks, though at higher prices.
The series of announcements confirms an important shift. During the early years of generative AI, focus was on who could train the largest model or achieve the highest scores. Now, the challenge is different: a lab may have a competitive model, yet lack the infrastructure to serve all its users effectively.
Kimi K3 offers one of the clearest examples. Just 48 hours after launching, Moonshot AI stopped accepting new subscribers. The company prioritized protecting service quality for existing clients rather than risking increased latency and instability due to growing demand.
Kimi K3 demonstrates the limits of serving a 2.8 trillion parameter model
Kimi K3 is a 2.8 trillion parameter model built on a Mixture of Experts (MoE) architecture. This design doesn’t use all parameters in each query; a routing system selects the most suitable experts to process each token, reducing the computational workload compared to a dense model of the same size.
However, that doesn’t mean it’s cheap or simple to deploy.
Moonshot recommends using supernodes with at least 64 accelerators to maintain high-bandwidth communication between experts. The full weights set would require around 1.5 TB in full precision, while a 4-bit quantized version would still need approximately 600 GB of memory.
The Kimi Delta Attention architecture significantly reduces the memory allocated for cached KV keys and values used to maintain context during generation. Yet, this improvement doesn’t eliminate the physical requirements of the model. The company still needs substantial high-bandwidth memory, interconnect capacity, and available GPUs for each request.
The launch caused daily sales and API consumption to spike, prompting Moonshot to split its subscription into two products. One covers the web app, work functions, and general Kimi service; the other, Kimi Code, targets programming tasks.
This separation has a technical basis. A typical conversation might last minutes, whereas a programming agent may work hours, analyze repositories, use tools, and generate millions of tokens. Treating both as a single workload hampers resource allocation.
Future release of the weights will also help Moonshot distribute costs. Cloud providers, large enterprises, and data center operators will be able to run Kimi K3 independently, although the model size keeps it out of reach for conventional servers.
DeepSeek V4 competes on price and efficiency in long context tasks
While Moonshot limited Kimi K3 sign-ups, DeepSeek began the full rollout of V4 Pro and V4 Flash, two models available in preview since April.
V4 Pro features 1.6 trillion total parameters and activates 49 billion per token. V4 Flash scales down to 284 billion total parameters and 13 billion active. Both provide a context window of one million tokens.
The architecture hasn’t changed significantly since the preview. Key improvements focus on reasoning, tool use, and agent behavior. Initial tests suggest fewer hallucinated responses and clearer answers during programming tasks.
DeepSeek also introduced an hourly pricing policy. Rates increase during peak demand hours in Beijing and decrease during off-peak times. This model, common in electricity, transportation, or cloud capacity markets, is rare among AI APIs so far.
Outside peak hours, V4 Pro costs around $0.87 per million tokens, while V4 Flash drops to about $0.28. These prices are much lower than those of high-end US models.
The comparison shouldn’t just consider cost per token; a cheaper model might produce longer responses, require more attempts, or generate errors, increasing overall costs. Nonetheless, the price difference is significant enough for many companies to consider DeepSeek for classification, data extraction, routine code generation, or large-scale automated processes.
Another notable change is the removal of the old identifiers deepseek-chat and deepseek-reasoner. Developers must migrate to the new V4 Pro and V4 Flash names to precisely control which variant they’re using.
Qwen 3.8 Max enters with 2.4 trillion parameters and multimodal support
Alibaba announced Qwen 3.8 Max Preview, a 2.4 trillion-parameter model using an MoE architecture. The company hasn’t disclosed how many parameters are active per token, which is essential for calculating its inference speed and cost precisely.
The model supports text, images, video, and documents. It features a context window close to 984,000 tokens and can generate responses up to 131,072 tokens long.
Alibaba has added adjustable reasoning effort levels—ranging from low to extra-high—to optimize resource use based on task difficulty, with reasoning enabled by default.
Qwen 3.8 Max is available via Alibaba’s token plans and through its Qoder and QoderWork platforms. The service supports OpenAI and Anthropic protocols, enabling many applications to switch endpoints without full re-integration.
Alibaba hasn’t published a standard price per million tokens or comprehensive benchmark results yet. Since this is a preview, details may change in upcoming weeks. The company has promised to release the weights but hasn’t confirmed the date or license.
Kimi, DeepSeek, and Qwen versus GPT-5.6 and Claude Opus 5
Chinese offerings come with low prices, open architectures or promises to open, and context windows near a million tokens. OpenAI and Anthropic maintain a different advantage: proprietary, more mature models, integrated tools, and infrastructures capable of higher service levels.
| Model | Architecture | Approximate Context | Starting Price | Availability |
|---|---|---|---|---|
| Kimi K3 | MoE, 2.8 trillion parameters | 1 million | $15 per million tokens | API, announced applications and weights |
| DeepSeek V4 Pro | MoE, 1.6 trillion total | 1 million | From $0.87 outside peak hours | API and open weights |
| DeepSeek V4 Flash | MoE, 284 billion total | 1 million | From $0.28 outside peak hours | API and open weights |
| Qwen 3.8 Max | MoE, 2.4 trillion parameters | Close to 1 million | No standard fee published | Closed preview |
| GPT-5.6 | Proprietary | Varies by version and mode | Up to $30 per million | OpenAI API and services |
| Claude Opus 5 | Proprietary | Up to 1 million | $25 per million | API and applications from Anthropic |
GPT-5.6 continues to excel in programming, research, design, and automation with tools. OpenAI offers variants and reasoning levels to balance speed, cost, and quality per project.
Claude Opus 5 especially competes in autonomous programming, repository analysis, extended agents, and computing. Anthropic positions it close to their largest model but at a lower task cost.
Both are more expensive than DeepSeek, though OpenAI and Anthropic benefit from larger scale—they control their infrastructure, models, tools, and many of the end-user interfaces.
Inference capacity becomes the new bottleneck
The Kimi K3 episode encapsulates the next phase of the tech race: creating advanced models isn’t enough; they must be served to millions without increasing latency, errors, or response costs.
China faces an added challenge: US restrictions limit access to the most advanced accelerators, forcing labs to maximize GPU efficiency and design more effective architectures.
These constraints have spurred advances in expert mixtures, efficient attention, quantization, and cached KV reduction. Yet, efficiency doesn’t fully replace physical capacity. Models with hundreds of billions of parameters still require large clusters, HBM memory, and high-speed interconnects between accelerators.
For developers, this means a broader ecosystem of competitive models tailored to different cost, privacy, openness, and performance needs. Companies can reserve GPT-5.6 or Claude Opus 5 for demanding tasks, while deploying DeepSeek, Kimi, or Qwen for routine or high-volume automation.
Deciding will shift from simply who is first in rankings to evaluating task completion costs, speed, token consumption, API stability, private deployment options, and actual infrastructure availability.
FAQs
Why did Kimi K3 stop accepting new subscribers?
Moonshot AI temporarily restricted sign-ups because demand was pushing their GPU infrastructure close to capacity limits. Existing customers continued to have access.
What is the difference between DeepSeek V4 Pro and V4 Flash?
V4 Pro activates more parameters and is designed for complex tasks. V4 Flash uses fewer resources, offering lower cost and latency for frequent workloads.
Is Qwen 3.8 Max an open model?
Alibaba announced intent to publish weights, but hasn’t specified the release date, licensing, or repository details yet.
Are Chinese models better than GPT-5.6 and Claude Opus 5?
It depends on the task. Chinese models excel in price, context, and openness, while GPT-5.6 and Claude Opus 5 have advantages in many professional workloads and offer more mature services.
