Update: Skip the GPUs — Kimi K3 is callable via API at $3/$15
The 2.8T open model that needed ~1.4TB to self-host now runs on managed endpoints — Together, Modal, OpenRouter, and Moonshot's own OpenAI-compatible API.
Copy markdown
The flip: managed hosting, not your rig
Yesterday's takeaway was that K3's 2.8T weights — roughly 1.4TB even in 4-bit — put self-hosting out of reach. The delta: day-0 managed inference from Together AI and Modal, plus an OpenRouter listing, so you call it like any other API.
The rate card: $3/$15, plus a 90% cache discount
Moonshot's platform lists $3.00 per million input tokens and $15.00 per million output, with cached input at $0.30 per million — a 90% cut that rewards reusing a long prompt instead of reloading it each call.
Drop-in: OpenAI- and Anthropic-compatible endpoints
The API speaks both the OpenAI and Anthropic wire formats, with tool calling, JSON-schema structured output, and context caching. Pointing an existing client at K3 is a base-URL-and-key change, not a rewrite.
1M context on a 2.8T MoE — 16 of 896 experts fire
You get the full 1,048,576-token window. The Stable LatentMoE design activates just 16 of 896 experts per token, which is how a 2.8-trillion-parameter model stays affordable to serve.
Build it this weekend — with one caveat
A repo-wide coding agent or whole-corpus doc pipeline on an open-weight base, no GPU purchase required. Caveat: third-party K3 prices aren't all officially confirmed yet — check each provider's live page before you point production spend at it.