Moonshot AI and Alibaba present new advanced open weight models that are only surpassed by Fable 5

Moonshot AI and Alibaba have unveiled Kimi K3 and Qwen3.8, respectively, two new AI models from open weights whose capabilities are second only to Fable 5.

In the artificial intelligence race, two Chinese companies have launched two new models based on open weights, that is, they are made available to developers and interested parties with their complete training weights so that they can work with them.

The first of them, Kimi K3, has been developed by Moonshot AI and has 2.8 billion parameters – the first to reach this size -, offers a context window of one million tokens and comes with native multimodal capabilities, including visuals.

The Beijing-based technology firm claims that it is its most capable model to date with performance that surpasses models such as GPT 5.5 and Claude Opus 4.8 in programming and agent tasks, although it is still below the most advanced Anthropic and OpenAI, Claude Fable 5 and GPT 5.6 Sol, as stated on its blog.

Specifically, Kimi K3 can work “in long engineering sessions, navigate massive repositories and coordinate terminal tools” with hardly any human intervention, and its visual intelligence capabilities allow it to optimize video game development from concepts, images and videos, alternating code with screenshots.

For its part, Alibaba has presented Qwen3.8, another large model that this time has 2.4 trillion parameters, which it claims is “one of the most powerful models available today, compatible with the main cutting-edge AI models and surpassed only by Fable 5”, as shared in X.

Kimi K3 open weights will be available from July 27, while Alibaba has indicated that Qwen3.8 will arrive “soon”, although Qwen3.8-Max-Preview is available on its test platform.

By launching their most advanced models as open weights, Chinese companies show a distribution opposite to that of American firms, which prioritize proprietary artificial intelligence and its distribution under more restrictive options.

Some of them have even joined together to combat attempts at adversarial distillation by Chinese companies, given the appearance of imitations of their AI products. In this sense, OpenAI openly accused DeepSeek of cloning American models, including its own.

Some components of Meta’s Llama models, distributed under an open source license, have been found in DeepSeek R1, which made the company rethink its strategy with the launch with Muse Spark, its first family of proprietary models.

Anthropic reported earlier this year that Chinese companies DeepSeek, Moonshot AI and MiniMax carried out distillation attacks to abuse its Claude model in order to take advantage of its capabilities and improve their models.