China’s Moonshot AI Releases Trillion-Parameter Model, Challenging US Dominance
Moonshot AI unveiled Kimi K2.5 on Monday, releasing what the Chinese startup calls “the most powerful open-source model to date” and intensifying competition in the race to democratize advanced artificial intelligence.
The model deploys 1 trillion parameters, using a mixture-of-experts architecture that activates 32 billion parameters per query. Available under a modified MIT license as a 595-gigabyte download on Hugging Face, K2.5 represents the latest salvo from China’s surging AI sector against established American players.
Performance metrics place K2.5 within striking distance of frontier models from OpenAI and Anthropic. On SWE-Bench Verified, a coding benchmark, the model achieved 76.8 percent accuracy compared to 80 percent for GPT-5.2 and 80.9 percent for Claude Opus 4.5. On certain tasks requiring tool use, K2.5 outperformed both rivals, scoring 50.2 on HLE-Full versus 45.5 for GPT and 43.2 for Claude.
The model’s distinguishing feature lies in what Moonshot terms Parallel-Agent Reinforcement Learning. This technique coordinates up to 100 sub-agents executing 1,500 tool calls simultaneously, reducing task completion time by 80 percent. Unlike conventional agent systems that assign predefined roles, K2.5’s agents operate autonomously, determining their own objectives and methods.
“At scale, the trade-off between vision and text capabilities disappears — they improve in unison,” Moonshot stated in its technical documentation. The model processes both video and text natively, trained on 15 trillion mixed tokens with a context window of 256,000 tokens.
Pricing undercuts Western competitors substantially. Moonshot charges 60 cents per million input tokens through its API, roughly 75 to 90 percent less than GPT-5 and Claude Sonnet 4.5. Average cost per task runs $5.20 versus $7.86 for Claude, according to company figures.
Yet the “open source” designation carries caveats. Running K2.5 requires at minimum 247 gigabytes of combined memory for a quantized version, or ideally 256 gigabytes of RAM paired with a 48-gigabyte GPU. The full model demands eight H200 graphics processors, hardware accessible primarily through cloud providers rather than individual researchers.
Reproducibility concerns have emerged since release. Community developers initially achieved only 20 percent accuracy on tool-call parsing using vLLM, an open-source inference engine, though this improved to 80 percent after adjustments. Moonshot created a “Kimi Vendor Verifier” tool to address discrepancies, maintaining that “results should be consistent” across implementations.
The release arrives as Moonshot pursues aggressive international expansion. The startup, valued at $4.8 billion with backing from Alibaba and Tencent, reported a fourfold increase in overseas API revenue between September and November 2025. An initial public offering may follow this year, according to people familiar with the matter.
Nathan Lambert, an AI researcher, suggested the model “puts pressure on incumbents, potentially prompting responses such as OpenAI accelerating GPT-5 or Anthropic pushing the envelope further.”
Whether K2.5 fulfills its promise depends partly on validation by independent researchers, who continue testing the model’s capabilities across diverse workloads. For enterprises evaluating AI infrastructure, the combination of competitive performance and lower costs presents a compelling alternative to established providers, provided they can navigate the substantial hardware requirements.
