Moonshot AI Opens Kimi K3 Weights, but Hardware Barriers Limit True Access
The 2.8 trillion-parameter model is technically open-source, but running it locally requires enterprise-grade infrastructure most developers cannot afford.
The Open-Weight Paradox
Moonshot AI released Kimi K3, a 2.8 trillion-parameter mixture-of-experts model, via API on July 16, 2026, with full open weights following on July 27. The release marks a technical milestone in open-model scaling while exposing a widening compute divide in the AI industry.
Despite the "open-weight" designation, Kimi K3's size creates prohibitive hardware requirements. Estimates for running the model at 4-bit quantization range from approximately 350GB of VRAM for weights alone to over 1.4TB when accounting for different quantization formats and KV cache overhead. This effectively restricts local deployment to elite research labs and large enterprises, pushing most developers toward API-based consumption.
Technical Specifications and Pricing
Kimi K3 supports a 1 million-token context window and processes both text and images. Moonshot AI positions the model for agentic coding workflows and long-horizon reasoning tasks.
API pricing is set at $3 per million input tokens and $15 per million output tokens, with cache hits priced at $0.30 per million tokens via providers like OpenRouter. This undercuts some proprietary competitors while maintaining premium positioning for frontier capabilities.
Performance and Rankings
On the public BenchLM leaderboard, Kimi K3 ranks fifth out of 215 models. A provisional subset previously showed the model at fourth place with a score of 80.96.
The Accessibility Question
Moonshot AI has pushed the upper bound of open-model sizes for nine of the last twelve months. Kimi K3 continues this trajectory, but the release highlights a fundamental tension in open-weight AI: publication does not guarantee practical accessibility.
The model's hardware requirements mean that "owning" Kimi K3 locally remains out of reach for most organizations, despite the weights being publicly available. This dynamic effectively recreates the API dependency that open-weight releases ostensibly aim to eliminate, raising questions about what "open" means in the era of trillion-parameter models.