After Kimi K3, Alibaba unveils Qwen 3.8, claims it's second only to Claude Fable 5

After weeks of US AI companies touting the cutting edge abilities of their AI models, it seems like Chinese companies don't want to be left behind either. Moonshot AI recently unveiled its Kimi K3 model that became so popular that the company had to pause new subscriptions in order to prioritize its current user base.
Now, Alibaba has also joined the race and has unveiled the new Qwen 3.8 model which it claims is only second to Anthropic's Fable 5.
The new AI model from Qwen comes with 2.4 trillion prameters and Alibaba has already take it available in preview to its Token Plan, Qoder and QoderWork users. However, the model will officially be launched soon and will also be going open weights, meaning that developers will be able to download, modify and deploy the model on their own infrastructure.
In a blogpost last week, Moonshot said that Kimi K3 is the first open source model to reach 2.8 trillion parameters. Notably, a higher parameter count allows a model to capture more patterns from its training data which can improve its reasoning, coding and language capabilities.
Moonshot says its latest AI model is adept at a wide range of tasks, including software engineering and long-horizon coding, scientific research, financial analysis, knowledge work, interactive presentations, image understanding, video editing and AI agents.
The company also said that K3 completed a number of complex engineering tasks autonomously, including optimising GPU kernels, building a Triton-like GPU compiler from scratch, creating a browser-based 3D game and even designing a chip for running smaller AI models.
Moonshot also claims the model can independently review dozens of research papers, generate thousands of lines of code and build interactive dashboards for scientific analysis.
In terms of comparisons with rival, Moonshot says that Kimi K3 outperforms GPT-5.5, Claude Opus 4.8 and GLM-5.2 across multiple software engineering and automation benchmarks while being close to Fable 5 and GPT-5.6 Sol on various coding related benchmarks.
Moonshot and Alibaba are also not the only Chinese startups to be launching cutting edge AI models that take on their American counterparts. Famously, DeepSeek had sent shockwaves across the Western markets last year with its V3 and R1 models but the company has not been able to replicate the same level of success with its newer models.
However, other Chinese startups have since take the lead. For instance, after the unveiling of Mythos 5 and Fable 5, Zhipu AI had claimed to have developed the GLM 5.2 model which could match the performance of Anthropic's powerful AI model on cybersecurity related tasks.
GLM-5.2 is a 753 billion-parameter model with a 1 million-token context window, which means that the model can be used to handle long coding sessions and large codebases without losing context.
On OpenRouter, GLM-5.2 ranked among the top 10 most-used AI models globally. Meanwhile, cybersecurity company Semgrep found that the Chinese AI model even outperformed Claude Opus 4.8 in certain tasks. Researchers also found that with additional prompting, GLM-5.2 and Claude Opus 4.8 could match Mythos' bug-finding abilities.
Aman Gupta is a Digital Content Producer at LiveMint with over 3.5 years of experience covering the technology landscape. He specializes in artificial intelligence and consumer technology, reporting on everything from the ethical debates around AI models to shifts in the smartphone market. <br> His reporting is grounded in first-hand testing, independent analysis, and a focus on how technology impacts everyday users. He holds a PG Diploma in Radio and Television Journalism from the Indian Institute of Mass Communication, Delhi (Class of 2022). <br> Outside the newsroom, he spends his time reading biographies, hunting for the perfect coffee beans, or planning his next trip. <br><br> You can find Aman on <a href="https://www.linkedin.com/in/aman-gupta-894180214">LinkedIn</a&gt and on X at <a href="https://x.com/nobugsfound">@nobugsfound</a&gt or reach him via email at <a href="aman.gupta@htdigital.in">aman.gupta@htdigital.in</a>.
Catch all the Business News, Market News, Breaking News Events and Latest News Updates on Live Mint. Download The Mint News App to get Daily Market Updates.
Oops! Looks like you have exceeded the limit to bookmark the image. Remove some to bookmark this image.

* Moonshot AI unveiled the Kimi K3 model.
* Alibaba unveiled the Qwen 3.8 model.
* The Qwen model has 2.4 trillion parameters.
* The Kimi K3 is the first open source model to reach 2.8 trillion parameters.
* Kimi K3 demonstrates capability in software engineering, coding, scientific research, and financial analysis.
* Kimi K3 autonomously completed tasks including optimizing GPU kernels and building a GPU compiler from scratch.
* Kimi K3 outperforms GPT-5.5, Claude Opus 4.8, and GLM-5.2 across multiple software engineering and automation benchmarks in some tests.
* Zhipu AI developed the GLM-5.2 model, which has 753 billion parameters and a 1 million-token context window.
* GLM-5.2 outperformed Claude Opus 4.8 in certain cybersecurity tasks according to Semgrep.

Chinese AI companies are competing with US firms in developing advanced large language models. Moonshot AI released the Kimi K3 model, which is noted as the first open source model to reach 2.8 trillion parameters, demonstrating capabilities in areas like software engineering, coding, and scientific research. Alibaba has followed this trend with its Qwen 3.8 model, claiming performance second only to Anthropic's Fable 5. The Qwen model has 2.4 trillion parameters and will be released as open weights for community use. Other Chinese entities, such as Zhipu AI, have also developed models like GLM-5.2, which achieved 753 billion parameters with a one million-token context window, demonstrating strong performance in cybersecurity tasks. These developments indicate an increasing pace of competition and innovation within the AI landscape outside of the US sphere.

The narrative presents an escalating technological race driven by parameter counts and demonstrated multi-modal reasoning capabilities, suggesting that competitive advantage is increasingly defined by scale and the depth of autonomous engineering tasks a model can perform rather than incremental feature releases. The comparison involving Kimi K3 against GPT-5.5 and Claude Opus suggests that proprietary performance benchmarks are evolving; the focus shifts from raw benchmark scores to the utility demonstrated in complex, real-world engineering workflows. The emergence of open weights models like Kimi K3, coupled with the performance metrics achieved by models like GLM-5.2 in specialized domains (like cybersecurity), points toward a decoupling between model size and domain-specific expertise, indicating that knowledge capture from massive training sets is becoming a more tangible asset than raw parameter count alone. This dynamic suggests a pattern where resource concentration in specific geographies fosters rapid, high-parameter innovation, forcing established Western leaders to rapidly re-evaluate their position against models emphasizing specialized functional competence alongside scale. What happens when performance metrics diverge significantly across different benchmarks? How does the open weights release affect the centralization of model control versus distributed innovation?

After Kimi K3, Alibaba unveils Qwen 3.8, claims it's second only to Claude Fable 5

Facts Only

Executive Summary

Full Take

After Kimi K3, Alibaba unveils Qwen 3.8, claims it's second only to Claude Fable 5

Facts Only

Executive Summary

Full Take