The 'Huobao' Team Tried Letting AI Be the Boss! Result: It Lost Money on Everything — Not Even Their Underwear Left

Kimi K2 API to be retired

IT Home (IT之家) has reported that Moon's Dark Side (月之暗面), the developer behind the open-source Kimi family of large models, announced the deprecation of the Kimi K2 series API. The notice says Kimi K2 will be taken offline on May 25, 2026 and will no longer be maintained or supported; users are advised to migrate to the newer Kimi model, kimi-k2.6, for continued support and "stronger multimodal understanding, reasoning, code and Agent capabilities."

Kimi K2 was first open-sourced in July last year as a trillion-parameter (1T) MoE architecture model with 32B active parameters, pitched at stronger code ability and general Agent tasks. In November the team released Kimi K2 Thinking, which it has been reported that the developers positioned as capable of "thinking while using tools" natively — reportedly performing up to 300 rounds of tool calls autonomously to tackle complex, multi-step problems.

Why it matters — cost, capability and claims

Maintaining trillion-parameter open models is expensive and operationally demanding, which helps explain why teams iterate and sunset earlier-generation APIs as newer versions roll out. For Western readers, note that China’s AI ecosystem has leaned heavily into open-source, domestically developed models both to accelerate innovation and to reduce reliance on Western toolchains amid export controls and geopolitical tensions. That context matters when evaluating fast product cycles and bold capability claims from Chinese labs.

The original item appears on Phoenix Net’s (凤凰网) social-media channel and carries the platform’s uploader disclaimer — the content was user-posted and the platform framed it as information storage. Separately, dramatic promotional claims — such as a so-called fully autonomous AI scientist "Robin" finishing months of human work in hours and discovering a blind-eye disease drug — remain extraordinary and, it has been reported that, are not independently verified. So: is handing the keys to an AI a futurist experiment or a headline-seeking risk? For now, most Chinese model teams are moving faster than regulators and users can fully test.