# Kimi K3 Puts Open AI Models on a Cost Test > Moonshot AI introduced Kimi K3, a 2.8T-parameter model with a 1M-token context window, live API access and weights promised by July 27. - Content type: NewsArticle - Section: News - Published: 2026-07-17T10:59:00.000Z - Publisher: Arkolith Newsroom - Canonical HTML: https://arkolith.com/news/news-kimi-k3-open-model - Plain Markdown: https://arkolith.com/news/news-kimi-k3-open-model.md - Topics: Kimi K3, Moonshot AI, Open models, AI infrastructure, AI coding ## Article Moonshot AI introduced Kimi K3 on July 16, putting a 2.8 trillion-parameter, 1 million-token-context model into public products and an API while promising full model weights by July 27. The launch gives developers a near-term test of whether an open 3T-class model can compete on useful agent work without matching the most powerful closed systems outright. The stakes are practical, not only technical. Kimi says the model is live in Kimi.com, Kimi Work, Kimi Code and the Kimi API. Its own blog also says overall performance still trails Claude Fable 5 and GPT 5.6 Sol, which makes the live question sharper: whether scale, long context and lower input prices are enough to change what builders choose for coding and knowledge-work agents. ## What Kimi released The [Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3) describes K3 as a native multimodal model built on Kimi Delta Attention and Attention Residuals, with 2.8 trillion parameters and a 1 million-token context window. It says the model activates 16 of 896 experts under a Stable LatentMoE framework and uses quantization-aware training for deployment across hardware. Kimi says the full model weights will be released by July 27, with a technical report due at the same time. That boundary matters. The model is usable now through hosted products and API access, but outside labs cannot yet inspect or run the full weights themselves. ## The cost test The [Kimi API Platform](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) lists the kimi-k3 model and shows three public API prices: $0.30 per million cache-hit input tokens, $3.00 per million cache-miss input tokens and $15.00 per million output tokens. Kimi's blog says its official API has a cache-hit rate above 90% in coding workloads, but that is a company-reported workload claim rather than an independent benchmark. That pricing puts the model in a different evaluation frame from a pure leaderboard story. Builders of long-running coding agents care about whether a model can keep context, use tools, avoid retry loops and stay affordable across many calls. The same question sits behind broader [agent-native API design](/blog/agent-native-apis-explained) and [source-citation workflows](/blog/how-ai-agents-cite-sources): long context helps, but repeated tool use and verifiable outputs still decide real reliability. ## What outside checks can and cannot prove The [Artificial Analysis model page](https://artificialanalysis.ai/models/kimi-k3) lists Kimi K3 as available and tracks independent model-performance, price and speed comparisons. It is useful corroboration that the model is now a public comparison target, but it does not settle Kimi's full benchmark claims or the upcoming open-weights release. Kimi's own benchmark table is unusually detailed, but it remains partly self-reported and uses different harnesses across tasks. The blog says some comparisons include fallback behavior for Claude Fable 5 and that more technical details will come with the report. A serious read should therefore separate three facts: Kimi K3 is live, the company is making large performance and price claims, and full independent validation is still incomplete. ## Why the timing matters The launch landed during a wider China AI-policy push. An [AP report](https://apnews.com/article/china-ai-tech-chips-xi-us-df4cfc7e1b260e765b5449b6d71a48e5) from the same World AI Conference described President Xi Jinping calling for global AI rules while criticizing restrictions that could widen access gaps. Kimi K3 is not itself a government policy, but it gives that argument a concrete commercial artifact: a Chinese lab trying to make a frontier-scale model broadly available. The next observable events are the July 27 weights release, the technical report, independent benchmark replication and whether inference partners can serve the model cheaply enough under real coding-agent load. If the weights slip, the technical report is thin, or independent tests underperform the launch claims, the story changes quickly. If they hold up, Kimi K3 becomes a live test of open-model economics rather than another benchmark screenshot. *This article is informational only and is not investment advice.* ## Citation and verification - Cite the canonical HTML article: https://arkolith.com/news/news-kimi-k3-open-model - Follow the links in the article to verify claims against the named primary or trustworthy sources. - Publication and update times above are the newsroom's recorded timestamps.