China’s AI race just produced two numbers that are hard to ignore: 2.4 trillion parameters and a price gap of more than 100X.

Alibaba on Monday introduced Qwen3.8-Max, its largest and most capable AI model yet, pushing the Chinese tech giant closer to the top of global AI rankings. Days earlier, DeepSeek released V4-Flash, a model priced so aggressively that benchmark testing puts its average cost at just three cents per test.

Alibaba is going after capability. DeepSeek is attacking cost. Together, the releases show how Chinese AI companies are trying to compete with OpenAI, Anthropic and Google without following the same economic playbook.

The launch comes just a week after another Chinese AI startup, Moonshot AI, released Kimi K3, a 2.8-trillion-parameter open-weight model and the largest free AI model released to date.

The back-to-back releases show how aggressively Chinese AI labs are scaling open models. Alibaba’s Qwen3.8-Max arrives with 2.4 trillion parameters, putting it close to Kimi K3’s 2.8 trillion, with both models capable of processing up to one million tokens at a time.

Alibaba shares jumped 7% in Hong Kong trading following the Qwen3.8-Max debut.

The competition is moving beyond who can build the smartest model. The next fight is over who can deliver enough intelligence at a price businesses and developers can afford to use at scale.

Alibaba unveils its largest AI model yet; DeepSeek’s latest model is ultra-low cost

Alibaba confirmed that Qwen3.8-Max will soon move beyond its initial release, with the company planning to make the model’s weights publicly available next week.

Announcing the release on X, Alibaba said: “Meet Qwen3.8-Max — our most capable model to date. Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!”

The announcement means Alibaba is opening two models at once: its flagship Qwen3.8-Max and the smaller Qwen3.8-27B. That gives developers a choice between Alibaba’s largest model and a much smaller option that should be easier to deploy with fewer computing resources.

Alibaba’s decision to release the weights puts Qwen3.8-Max alongside a growing group of Chinese models that developers can download and run themselves. DeepSeek has built much of its global following around that approach, and Moonshot AI joined the trillion-parameter race with Kimi K3 just a week earlier.

The competition with U.S. AI companies is no longer centered solely on which lab produces the highest benchmark score. Chinese AI companies are pairing increasingly capable models with open weights and sharply lower prices, creating a different route to winning developers.

Qwen3.8-Max’s 2.4 trillion parameters make it one of the largest AI models released by a Chinese company. Its mixture-of-experts architecture means only a fraction of those parameters need to be active for each request, an approach aimed at keeping inference costs and response times manageable as models grow.

Parameter counts offer an imperfect measure of AI capability. More parameters do not automatically translate into a better model. They do give some indication of the scale of the system and the computing and training effort behind it.

Qwen3.8-Max made an immediate showing on Arena.AI, the crowdsourced platform where users compare AI models without initially knowing which model generated each response.

The model became the highest-ranked Chinese text model on the platform soon after appearing. It still trails Anthropic’s Claude Fable 5 and three Claude Opus variants.

Alibaba fared even better in multimodal testing.

Qwen3.8-Max climbed to second place globally among models evaluated on images and other visual material, behind a Claude Fable 5 variant.

The result puts Alibaba in increasingly close competition with some of the strongest models coming out of U.S. AI labs.

One million tokens and a massive MoE architecture

Qwen3.8-Max can work across text, images, and video and process as many as one million tokens in a single context window.

Moonshot AI’s Kimi K3 offers similar capabilities.

A million-token context window lets a model process unusually large amounts of information in one session. That could mean hundreds of pages of documents, long legal files, or substantial portions of a software repository.

Alibaba says Qwen3.8-Max completed a software-engineering project over a 16-day period, pointing to the growing effort across the AI industry to build models capable of carrying out longer-running tasks rather than answering isolated prompts.

Qwen3.8-Max (Click to enlarge)

The size of Qwen3.8-Max tells only part of the story.

Alibaba built the model using a mixture-of-experts, or MoE, architecture. Instead of activating all 2.4 trillion parameters for every request, the system calls on specialized portions of the model depending on the task.

Only about 95 billion parameters are active at a given time.

That distinction matters economically. Running all 2.4 trillion parameters for every request would require enormous computing resources. Activating a fraction of the model gives Alibaba a way to build at trillion-parameter scale without paying the full inference cost on every query.

The architecture reflects one of the central engineering problems facing AI companies: bigger models are useful only if companies can afford to run them.

DeepSeek attacks the other side of the equation

DeepSeek is pushing that argument much further.

Its V4-Flash model, released Friday, is the cheapest major model to run across benchmark tests tracked by San Francisco AI research firm Artificial Analysis.

DeepSeek charges $0.14 per million input tokens and $0.28 per million output tokens.

The headline token prices are already low. The benchmark economics are more striking.

Artificial Analysis estimates that V4-Flash costs an average of just $0.03 per test.

Kimi K3 averages $0.86.

OpenAI’s GPT-5.6 Sol comes in at $1.86.

Anthropic’s Claude Fable 5 costs $3.15.

That puts Claude Fable 5 at roughly 105 times the average benchmark cost of DeepSeek V4-Flash.

The comparison matters more than token prices alone. A cheap model can still become expensive if it needs far more tokens or repeated attempts to finish the same task. Artificial Analysis measures the cost of completing benchmark workloads, giving developers a better view of what a model may cost in actual use.

That three-cent figure could become one of DeepSeek’s strongest selling points.

For companies processing millions of AI requests, small differences in inference costs can turn into substantial infrastructure bills. A model does not necessarily need to beat the best system on every benchmark if it can perform the required job at a fraction of the cost.

China’s open-weight strategy is becoming clearer

Alibaba and DeepSeek share another trait that separates much of China’s AI industry from its biggest American competitors.

Both are backing open-weight models.

Open-weight releases give developers access to the learned model parameters, allowing them to download, run and modify models on their own infrastructure.

OpenAI, Anthropic and Google have centered their flagship businesses around closed models accessed through APIs and hosted services.

That creates two competing approaches to AI distribution.

One asks developers to access intelligence through infrastructure controlled by the model provider. The other gives developers greater control over where models run, how they are modified and what infrastructure supports them.

Chinese AI companies appear increasingly willing to use openness and price as distribution tools.

“Chinese AI companies have found an important market. Many business workflows do not need the industry’s very best model,” Lian Jye Su, chief analyst at research firm Omdia, told Reuters.

“They need models that are good enough, affordable, transparent and accessible, and open-weight models ⁠help meet that demand.”

That may prove to be the larger story behind this generation of Chinese models.

DeepSeek is repeating a familiar move

DeepSeek has been here before.

The startup became one of the biggest AI stories of early 2025 after its R1 and V3 models challenged assumptions about how much money and computing infrastructure were required to build competitive AI systems.

The releases helped trigger a global technology stock selloff and forced investors to question whether the enormous AI infrastructure budgets of U.S. technology companies would translate into lasting advantages.

V4-Flash pushes the same argument from another direction.

Rather than simply asking whether Chinese labs can build models competitive with American systems, the new question is how low the cost of useful AI can go.

DeepSeek, which sources have said is preparing for a potential IPO, has strong incentives to keep pressing that advantage.

The company does not need to win every benchmark to create pressure across the industry. It needs to make developers question why they should pay dramatically more for workloads that a cheaper model can handle.

The AI race is becoming a price war

For the past several years, AI competition has largely been described through benchmark scores, model sizes, GPU clusters and training budgets.

Alibaba and DeepSeek are adding another metric: cost per useful result.

That shift could matter more to businesses than another few points on a leaderboard.

Companies deploying AI across customer service, software development, document processing, research and internal operations eventually have to turn model performance into an operating expense. The economics become harder to ignore once millions or billions of tokens start moving through production systems.

Alibaba’s Qwen3.8-Max suggests Chinese labs are still willing to build at enormous scale. DeepSeek V4-Flash shows they are equally interested in driving the price of inference down.

Those strategies are converging around the same idea.

The winning AI model may not always be the smartest model available. For a large share of commercial workloads, it could be the model that is smart enough and cheap enough to run everywhere.