What Kimi K3 Is (a New 2.8-Trillion-Parameter Model)
Kimi K3 is a large language model announced on July 16, 2026 by Moonshot AI, an AI company based in Beijing, China. It can be used on the website and app of the "Kimi" chat service, and a developer API launched at the same time. Its biggest distinction is that Moonshot AI has promised to publicly release the model itself (the weights)—one of the largest ever at a total of 2.8 trillion (2.8T) parameters—by July 27, 2026.
Key specifications of Kimi K3 (official announcement)
Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window. — from the opening of the announcement (2.8T parameters, native vision, 1M context)
An MoE Design That "Uses Parts of" 2.8 Trillion Parameters
Kimi K3 is an MoE (Mixture of Experts) model. MoE refers to a design that holds many "expert" components inside the model and activates only some of them depending on the input. In Kimi K3, only 16 of 896 experts are selected and run, so the amount of computation per query stays modest relative to the sheer size of the total parameter count. The key to the design is that rather than running the full 2.8 trillion parameters every time, it selectively uses only the parts it needs, keeping speed and cost practical.
effectively activating 16 out of 896 experts when paired with a Stable LatentMoE framework — from the description of the expert configuration
Support for 1M-Token Long Text and Image Input
The context (the amount of text it can read at once) is up to one million tokens (a token is the unit an AI uses to count segments of text), making it well suited to handing over long documents or codebases all at once. It is also designed from the start to accept image input as "native vision," rather than adding image support later. The everyday pattern of showing a chart or screenshot while giving instructions in text can be handled by a single model.
Kimi K3's Performance (Comparison with Claude)
What drew the most attention in the announcement were the performance claims against the top U.S. models. The main points are as follows.
Summary of Kimi K3 evaluations (official claims and external reports)
Moonshot AI Claims It Beats Opus 4.8 and Competes with Fable 5
Moonshot AI says that in coding-task evaluations, Kimi K3 substantially outperformed Anthropic's Claude Opus 4.8 and competed with the top-tier Claude Fable 5. The very fact that a China-born model directly claims parity with the top U.S. models at the time of release is the big news in this announcement. That said, these are figures from the developer itself. Third-party verification is still to come.
Kimi K3 performed competitively with Fable 5 (with fallback) and substantially outperformed Opus 4.8 — from the kernel-optimization (coding) evaluation
External Evaluations Also Report Frontend First Place and Overall Second
It is not only the official claims. Simon Willison, a UK-based AI researcher, reports that Kimi K3 overtook Claude Fable 5 to take first place on Arena.ai's frontend coding comparison, which pits models against each other by user votes. Also, on the research firm Artificial Analysis's long-horizon knowledge-work evaluation, it scored an Elo (a rating for strength in head-to-head formats) of 1547, second behind Fable 5. "First in some of its strong areas, and just behind the very top overall"—that is how Kimi K3 stands based on external evaluations right after the announcement.
The model is also now the leading model on Arena.ai’s Frontend Code arena, surpassing even Claude Fable 5. — from the note on the Arena.ai frontend comparison
On our private long-horizon knowledge work evaluation, Kimi K3 reaches an overall Elo of 1547, +732 points from Kimi K2.6 and behind only Claude Fable 5. — from the quoted Artificial Analysis tally
Kimi K3's Pricing and Planned Open-Weight Release
Alongside performance, cost and how to obtain it are natural concerns. The current form of availability and the promised weights release are as follows.
How to use Kimi K3 and its cost (as of the announcement)
API Pricing Drops Sharply with Caching
The API is priced at $3 per million input tokens and $15 per million output tokens. When the same document is read repeatedly, input that hits the cache (reuse of content already read) drops to $0.30. In workflows that go back and forth many times on top of a long document, that difference has a real effect on cost.
Pricing is $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. — from the availability and pricing description
The Full Weights Are Planned for Release by July 27, 2026
At the time of the announcement, only the API and chat are distributed, but Moonshot AI clearly states it will release the full model weights by July 27, 2026. If realized, it will be one of the largest open-weight models ever, far exceeding GLM-5.2. That said, a model of 2.8 trillion parameters is not something you can run on a personal computer, and for many users API or chat access will remain the realistic entry point even after release. For choosing how to try open models in your own environment, see also practical coding with local LLMs.
The full model weights will be released by July 27, 2026. — from the note on the planned weights release
When researching overseas models like Kimi K3, you increasingly need to read through English official blogs and model cards. If you want an AI to summarize long English material, preparing it in Markdown format first preserves the structure of headings and tables and improves accuracy. Pasting a web page as-is tends to mix in styling information, so tidying the source before handing it over is the shortcut to stable results.
Kimi K3 is a model that simultaneously puts forward one of the largest-ever scales at 2.8 trillion parameters, a claim of standing alongside the top models, and a promise to publicly release its weights. Third-party verification is still ahead, but if the weights are released as promised by July 27, it will be a milestone in the open-model landscape. A sound approach is to first check the quality of its responses in chat, estimate the cache-based API costs for development uses, and then judge together with the follow-up reports after release.
