Alibaba has unveiled Qwen3.8-Max, its largest and most capable artificial intelligence model to date, built with 2.4 trillion parameters. The model was previewed on 19 July 2026 at the World Artificial Intelligence Conference (WAIC) in Shanghai, China. It trails Moonshot AI’s Kimi K3, a competing Chinese model with 2.8 trillion parameters, in raw size.
What Is Qwen3.8-Max?
Qwen3.8-Max is the new flagship of the Qwen family, the series of large language models (LLMs) developed by Alibaba Cloud, the cloud computing arm of the Chinese e-commerce giant Alibaba Group. A large language model is an AI system trained on enormous volumes of text to understand and generate human-like language.
A parameter is a numerical value that a model adjusts during training, and it captures what the model has learned from data. In simple terms, the more parameters a model has, the larger its memory and the more patterns it can recognise. The count of 2.4 trillion makes Qwen3.8-Max Alibaba’s biggest model yet and its first multimodal model above 1 trillion parameters.
The model is multimodal, meaning it can process text, images, video, and documents together in a single system rather than handling each format separately. It carries a 1 million token context window, the amount of text a model can hold in memory at once, which is roughly equal to eight hundred thousand words or several full-length novels.
Qwen3.8-Max is built on a Mixture of Experts (MoE) architecture. In this design, the model is divided into many smaller specialist networks, or “experts”, and only a fraction of them are activated for any single query. This keeps the cost of running a 2.4 trillion-parameter model manageable, because the whole network does not wake up for every question. Alibaba has not disclosed how many parameters are active per query.
Alibaba says the model outperforms its predecessor, Qwen3.7-Max, in coding, full-stack development, data analysis, and office workflows. The company has described it as “one of the most powerful models available today, comparable to leading frontier systems”, a category that includes models from American labs such as OpenAI and Anthropic.
A Rivalry Built on Trillions of Parameters
The Qwen3.8-Max preview landed barely two days after Moonshot AI, a Beijing-based startup, released Kimi K3 in mid-July 2026. Kimi K3 carries 2.8 trillion parameters, making it both China’s largest AI model to date and the world’s largest open-weight model, one whose internal weights, the trained parameters of the network, can be freely downloaded, modified, and run by developers.
Moonshot AI says Kimi K3 uses a Mixture of Experts design that activates just 16 of its 896 experts per token, roughly 1.8 percent of the pool, and it supports a 1 million token context window with native vision, meaning it can interpret images. Independent evaluations from platforms such as Artificial Analysis and Arena.ai place the model on a par with leading American systems, and it ranked first in a blind frontend-coding test ahead of Anthropic’s Claude Fable 5.
| Model | Developer | Parameters | Released |
|---|---|---|---|
| Qwen3.8-Max | Alibaba | 2.4 trillion | July 2026 (preview) |
| Kimi K3 | Moonshot AI | 2.8 trillion | July 2026 |
| Qwen3.7-Max | Alibaba | more than 1 trillion | May 2026 |
| Qwen3-Max | Alibaba | more than 1 trillion | September 2025 |
The table shows how quickly the race has moved. Alibaba’s previous flagship, Qwen3.7-Max, released in May 2026, already exceeded 1 trillion parameters, and Qwen3.8-Max now more than doubles that figure within two months. The two launches together signal that Chinese labs have shifted the centre of competition in AI from thousands of crores of rupees in funding alone to the sheer scale of the models being trained.
The Qwen Family: From Tongyi Qianwen to Trillion-Parameter Flagships
The Qwen series began in April 2023 under the name Tongyi Qianwen, which translates as “to comprehend the meaning and answer a thousand kinds of questions”. Alibaba opened the model to public use in September of that year, after regulatory clearance from Chinese authorities, and its early versions were built on the Llama architecture released by the American company Meta.
Through 2024 and 2025, the family expanded steadily. Qwen2 arrived in June 2024, and in September 2024 Alibaba released more than 100 open-source models at once, with many licensed under the permissive Apache 2.0 license, which allows anyone to use, modify, and distribute the code freely. In January 2025 the company launched Qwen2.5-Max, and on 28 April 2025 it unveiled the Qwen3 family, which included both dense models and MoE variants trained on 36 trillion tokens in 119 languages.
Alibaba has followed a dual strategy. Many Qwen models, especially the smaller ones, are released as open source and have been downloaded more than 40 million times, spawning over 200,000 variations by third-party developers on the Hugging Face platform. The largest flagship models, including Qwen3-Max and now Qwen3.8-Max, have historically remained proprietary, served only through Alibaba Cloud. For Qwen3.8-Max, Alibaba has said it will open-source the model’s weights for the first time at the flagship level, though no date, license, or repository has been confirmed yet.
The Qwen team operates under the Tongyi Laboratory, which Alibaba restructured in 2026 into a dedicated unit focused on developing the Qwen models. The lab is part of Alibaba’s broader push into artificial intelligence, alongside the company’s Alibaba Token Hub, a business group created in March 2026 to oversee AI-related work.
Moonshot AI: The Startup Behind Kimi
Moonshot AI was founded in March 2023 in Beijing by Yang Zhilin, a graduate of Tsinghua University. The company is backed by major Chinese technology firms, including Alibaba and Tencent, and it raised $2 billion in a funding round in May 2026 that valued the startup at more than $20 billion.
Moonshot is best known for Kimi, a chatbot that became popular in China for handling very long documents, and it has quickly risen to the front of China’s generative AI ecosystem. Kimi K3, the company’s flagship, is notable not just for its size but for its open-weight release, with full weights due by 27 July 2026. This makes it the first model in the three-trillion-parameter class that anyone can download and run on their own hardware.
The announcement had an immediate market effect. Shares of Moonshot’s Chinese rivals, including Zhipu AI and MiniMax, fell sharply in Hong Kong, by about 28 percent and 16 percent respectively, on the day Kimi K3 was revealed. The reaction mirrored the global shock that followed DeepSeek-R1 in January 2025, the Chinese reasoning model whose low-cost development triggered a sharp fall in technology stocks and was widely compared to a “Sputnik moment” for the AI industry.
What the Model Size Race Really Means
The headline numbers conceal an important technical reality: total parameters do not directly decide how powerful a model feels in practice. In a Mixture of Experts design, only a small fraction of the network runs on each query, so the figure of 2.4 trillion says little about speed or running cost. Alibaba’s earlier Qwen3-235B-A22B model, for example, has 235 billion total parameters but activates only 22 billion per token, and the company has not yet disclosed the active-parameter count for Qwen3.8-Max.
There is also a question of verification. Alibaba has described Qwen3.8-Max as “second only to Fable 5” among the systems it benchmarked, but it has not published a full benchmark table, a model card detailing the model’s limits, or an independent evaluation. The performance claims rest on the company’s internal testing until third parties score the model.
What is verifiable is the broader trend. Chinese AI labs are producing frontier-scale models despite American export controls that restrict their access to Nvidia’s most advanced chips. China has responded with architectural innovation, including efficient MoE designs, hybrid attention mechanisms such as Kimi Delta Attention, and reliance on domestic hardware from companies like Huawei. Kimi K3, for instance, runs efficiently on Nvidia H200 chips and on an unnamed alternative graphics processor, according to the company.
The practical stakes extend beyond China. Cheaper, open-weight Chinese models are already gaining adoption among developers worldwide, because they close the performance gap with American rivals while costing far less. The US government has taken note; in June 2026 it ordered Anthropic to suspend access to its Fable 5 and Mythos 5 models over cybersecurity concerns, and lawmakers have debated whether American companies should use Chinese open-source models at all.
India’s AI Position in a Trillion-Parameter World
The escalating race between American and Chinese labs frames India’s own push for AI self-reliance. The government launched the IndiaAI Mission in March 2024 with an outlay of ₹10,372 crore to build a national AI ecosystem, including state-of-the-art compute infrastructure under a public-private partnership.
IndiaAI has provisioned more than 38,000 GPUs so far, with a plan to scale sovereign compute capacity toward 58,000 GPUs, covering roughly 40 percent of the compute cost for eligible users such as startups and researchers. The IndiaAI Compute Capacity initiative is designed to give Indian innovators affordable access to the computing power that training and running frontier-scale models requires, at up to 40 percent reduced cost.
Building a 2.4 trillion-parameter model demands compute resources on a scale few countries can match, and India’s early focus on compute access, data, and foundational models through the mission reflects a bet that affordable infrastructure, rather than sheer parameter counts, will determine who benefits from the coming wave of AI applications.
The Way Forward
The rivalry between Alibaba’s Qwen3.8-Max and Moonshot’s Kimi K3 shows that the frontier of AI is being pushed from two directions: the parameter count, which keeps climbing, and the cost of running these giants, which keeps falling through efficient architecture. Both models are expected to be followed by smaller, distilled versions that ordinary developers can actually deploy.
For Alibaba, the challenge is whether it can match the open-weight momentum that Kimi K3 has created. Moonshot has already shipped its weights, while Alibaba has only promised them. For the wider industry, the race means cheaper models, faster innovation, and a steady erosion of the assumption that only American firms can lead frontier AI. The next milestone will come when third-party benchmarks verify, or dispute, the claims that both companies have made.
Key Takeaways
- Qwen3.8-Max, unveiled on 19 July 2026 at the World AI Conference (WAIC) in Shanghai, is Alibaba’s largest AI model with 2.4 trillion parameters.
- Kimi K3 from Moonshot AI holds 2.8 trillion parameters, making it China’s largest model and the world’s largest open-weight model.
- Qwen is the AI model family of Alibaba Cloud, launched in April 2023 under the name Tongyi Qianwen.
- Moonshot AI, founded in March 2023 in Beijing by Yang Zhilin, is backed by Alibaba and Tencent and valued at more than $20 billion.
- Qwen3.8-Max is multimodal, with a 1 million token context window and a Mixture of Experts (MoE) architecture.
- IndiaAI Mission, launched in March 2024 with an outlay of ₹10,372 crore, is India’s national initiative to build AI compute and model infrastructure.
- The launches follow American export controls on advanced chips, which Chinese labs have responded to with efficient MoE and attention architectures.
