Moonshot AI Releases 2.78-Trillion-Parameter Open Model at Lower Cost

What Happened

Chinese AI company Moonshot AI has released Kimi K3, an open-weight large language model with 2.78 trillion parameters, the company announced. The model scores within a few points of leading proprietary systems on standard benchmarks, according to reporting by BigGo Finance, while being developed at a cost substantially below what frontier proprietary labs have historically spent on comparable models.

What the Model Is

Kimi K3 is an open-weight model, meaning its parameters are made publicly available for external use, inspection, and deployment. The 2.78-trillion-parameter scale places it among the largest openly released models to date. According to the wire report, Moonshot AI's benchmark results position Kimi K3 within a few points of top proprietary systems, which include models from OpenAI, Google, and Anthropic that are not publicly released and are accessible only through commercial APIs.

The model's release continues a pattern in which Chinese AI laboratories have published large-scale open-weight models that perform competitively against closed Western counterparts. Earlier in 2025, DeepSeek's R1 model drew significant industry attention for achieving strong benchmark performance at reported training costs far below those disclosed by major US laboratories.

The Cost Question

The central claim accompanying Kimi K3's release is that frontier-level benchmark performance no longer requires frontier-level capital expenditure. Moonshot AI has not publicly disclosed specific training costs in the available wire reports, and independent verification of the cost claims has not been reported. The characterisation of reduced spending relative to comparable models is based on reporting from BigGo Finance citing the company's own positioning.

The broader context is significant. US AI laboratories including OpenAI, Google DeepMind, and Anthropic have each disclosed or implied training runs costing hundreds of millions to billions of dollars for their most capable models. If Kimi K3's performance claims hold up under independent evaluation, its release would represent a continued compression of the cost curve for large-scale AI model development.

About Moonshot AI

Moonshot AI is a Beijing-based artificial intelligence company founded in 2023. The company's Kimi product line has been its primary public-facing AI offering, targeting both consumer and developer audiences. Moonshot AI raised substantial venture capital funding in 2024, with reported valuations placing it among China's most prominent AI startups. Kimi K3 represents the company's most technically ambitious release to date based on publicly available information.

Open-Weight Models and the Competitive Landscape

The release of large open-weight models has become a distinct competitive axis in the AI industry, separate from the closed-model ecosystems operated by major US companies. Meta's Llama series established a precedent for large-scale open releases from a well-capitalised company. Chinese laboratories including DeepSeek, Alibaba's Qwen team, and now Moonshot AI have each contributed open-weight models that benchmark analyses have placed close to or within range of closed proprietary systems.

Open-weight releases allow developers, researchers, and enterprises to run models locally, fine-tune them for specific applications, and avoid per-token API costs. This has made open-weight model releases a significant factor in enterprise AI procurement decisions, particularly for organisations with data privacy requirements or cost sensitivity at scale.

What Happens Next

Independent benchmark evaluations of Kimi K3 by third-party researchers and AI testing organisations are expected to follow the public release, with results likely to be published in the coming weeks across platforms including Hugging Face's Open LLM Leaderboard and similar evaluation frameworks.