Kimi K3: A 2.8 Trillion-Parameter Model With a Million-Token Window Available on Microsoft Foundry

Moonshot AI's Kimi K3, an open-weight model with 2.8 trillion parameters and a 1 million-token context window, is available for deployment on Microsoft Foundry through Fireworks AI. The model is hosted on enterprise deployment infrastructure.

The context window is 1 million tokens, roughly 750,000 words—enough to cover about ten novels or an entire codebase in a single request. This eliminates the need for chunking or retrieval-augmented generation when working with long documents, meeting transcripts, or large code repositories.

Fireworks AI is providing the optimized inference layer, while Microsoft Foundry handles the deployment, governance, and data zone isolation that enterprise customers typically require. Kimi K3 is not running on public APIs in this setup; it lives inside Foundry's managed environment.

The pricing is structured as follows: $3.30 per million input tokens, $16.50 per million output tokens, and $0.33 per million cached input tokens. That output price is relatively high compared to many commonly deployed models, which makes the cost-per-task sensitive to how much the model generates per query.

Uses for a 1M-token window include passing an entire Node.js application—package.json, source files, dependencies—into a single prompt and asking Kimi K3 to identify unused imports, potential SQL injection points, or architectural improvements. The same approach applies to legal contracts, research papers, or compliance documents. The model sees everything at once rather than reconstructing context across fragments.

The source presents these use cases as examples rather than benchmarked results. The model may be able to process 750,000 words in a single pass, but that does not guarantee reliable retrieval of obscure details buried in the middle of that input. Long-context models have improved significantly, but attention distribution across a million tokens remains an active area of research. Additionally, 2.8 trillion parameters is an enormous model, and inference cost and latency at that scale are not trivial—Fireworks' pricing reflects this.

Customers interested in deploying Kimi K3 through Foundry will need to check applicable availability, terms, and pricing with both Microsoft and Fireworks directly. The model is open-weight, but the deployment path runs through a managed service rather than a self-hosted option.