MANY MINDEDA 300-person Chinese lab builds a frontier-class model without the usual GPU stockpile
Moonshot AI, a Chinese startup of roughly 300 people, has released an AI model called Kimi K3. Early assessments place it on par with Anthropic's Opus 4.8, though below the leading frontier systems. Notably, the team built its own Mooncake training stack partly because it lacked sufficient GPUs — a constraint that would normally rule out competing at this level.
The model is large, reportedly 2.8 trillion parameters according to SemiAnalysis, too big to fit on a single Nvidia DGX B200 even with FP4 quantization. But on cost it undercuts rivals: Artificial Analysis puts it at an average $0.94 per task, against $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. SemiAnalysis had recently argued Chinese labs were too compute-poor to reach the frontier; Kimi K3 complicates that claim. Dylan Patel of SemiAnalysis noted Chinese firms can rent GPUs outside China, blunting part of export controls.
When a small team matches expensive models at lower cost and releases openly, capable AI stops being the property of a few well-funded labs. Cheaper inference and open weights push a knowledge and communication tool toward something closer to a shared utility.
The picture is still fuzzy. The gap to top models is described as unclear, and the assessments rest on early tests and individual expert opinions. One reviewer called it token-hungry; another, from OpenAI, praised it while having an obvious stake in the closed-versus-open debate. Comparisons will firm up as more people run it.
Source: The Decoder