Kimi K3 Ships 2.8T Parameters

TL;DR

- Kimi K3 launched July 16, 2026 as a 2.8-trillion-parameter open-weight model from Moonshot AI, more than double the roughly 1T parameters in the previous Kimi K2.6. - API pricing runs $3 per million input tokens and $15 per million output tokens, with cached input at a competitive rate designed to undercut most Western flagships. - Moonshot claims performance on par with Claude Opus 4.8 and GPT-5.5, though K3 still trails Claude Fable 5 and GPT 5.6 Sol. - The model ships with a 1-million-token context window built for long-horizon software engineering, complex reasoning, and knowledge work. - Independent benchmarks are still pending, but one scoring source already rates K3 at 79/100 composite in coding. If you pay premium API rates for coding tasks, test this endpoint.

Moonshot AI released Kimi K3 on July 16, 2026, and the specs alone force a conversation about API pricing. This is a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window, released as an open-weight model with live API documentation. Moonshot calls it their largest and strongest base model to date, positioned for long-horizon software engineering, complex reasoning, and knowledge work. The company claims Kimi K3 matches Claude Opus 4.8 and GPT-5.5 on overall performance.

Those are fighting words at $3 per million input tokens.

What is Kimi K3 and how large is it?

Kimi K3 is a 2.8-trillion-parameter large language model built on a Mixture-of-Experts architecture.

That parameter count makes it the world's largest open model by parameter count, according to Moonshot AI. And it more than doubles the roughly 1 trillion parameters in the previous Kimi K2.6.

The model launched simultaneously through the Kimi app, Playground, and API under the model ID `kimi-k3`.

API documentation went live at release, but independent benchmarks are still catching up to the launch. What Moonshot has shared so far: official results showing competitive performance on SWE-bench Verified, LiveCodeBench, Tau2, and AIME. The company says K3 reached a leading level among current open models on those benchmarks and partially approached top-tier closed models on specific tasks.

One independent scoring source, LM Market Cap, already places Kimi K3 at #61 in the Coding category with a composite score of 79/100. That ranking will shift as more evaluators test the model in the coming weeks. But it gives you a concrete data point on day one rather than marketing claims alone.

How much does Kimi K3 cost compared to alternatives?

The pricing is where things get uncomfortable for Western API providers. Kimi K3 costs $3 per million non-cached input tokens and $15 per million output tokens. Moonshot explicitly designed this to undercut most Western flagships.

For small agencies and solo developers, the math matters immediately.

If your workload is context-heavy with relatively short outputs (code review, document analysis, classification, summarization), your effective cost per query drops hard. The cached input rate rewards repeated context window usage. That is exactly the pattern you see in multi-step agent workflows and repository-scale code analysis.

The output pricing at $15/M is the steeper half of the equation. If you are generating large code files or long-form documents, you pay the full output rate on every token produced.

That number is still competitive against models in the same performance tier. But it is the line item to model carefully in your cost projections.

My agency routes API traffic across multiple providers depending on the task.

Adding Kimi K3 as a cost-optimization layer for coding and analysis workloads is the obvious first test. You are not ripping out infrastructure or migrating pipelines. You are adding an endpoint and comparing the bill.

Can Kimi K3 actually compete with Claude and GPT?

Moonshot AI said Kimi K3 is on par with Anthropic's Claude Opus 4.8 and OpenAI GPT-5.5. That is a company claim, not an independent verdict.

The same source acknowledges K3 still trails the most powerful proprietary models, including Claude Fable 5 and GPT 5.6 Sol.

So the honest read is: Kimi K3 appears competitive with mid-tier frontier models from Anthropic and OpenAI, not the absolute top of the stack. And it may lead the open-weight category outright.

The competitive results on SWE-bench Verified are the number I care about most.

That benchmark measures real software engineering tasks: bug fixes, feature implementation, actual codebase navigation. If Kimi K3 performs there the way the official numbers suggest, it is a legitimate coding model. Something you would trust on a pull request, not just a chatbot that happens to string together Python snippets.

The open-weight angle changes the strategic picture too. Moonshot is releasing K3 as an open-weight model, which means developers can download and modify it. A 2.8-trillion-parameter open model is not something you self-host on a laptop. But for teams with real hardware or cloud budgets, the per-token economics shift dramatically when you own the weights instead of renting them.

What should small teams do with Kimi K3 today?

Start with the API.

The pricing is low enough that you can run a real workload test for under $20. Pick a task you currently send to Claude or GPT, something with a measurable quality bar. And route it through the `kimi-k3` endpoint. Compare output quality and cost side by side.

If you build agent workflows that need long context windows, the 1-million-token limit is a genuine advantage. Multi-step coding agents, repository-scale analysis, sustained tool use across large codebases. Those are the workloads Moonshot built K3 to handle. The model is positioned for long-horizon software engineering, which in practice means jobs that take hours, not prompts that take seconds.

For teams considering self-hosting down the road: watch for the open-weight release.

The economics of running a 2.8T-parameter model only work at scale, but at scale, they work very well. If Moonshot delivers full weights and the benchmark claims hold under independent testing, Kimi K3 becomes the default open-weight option for anyone who needs frontier-tier coding capacity without per-token lock-in.

Moonshot shipped frontier-tier model capacity at disruption pricing today. The benchmark claims need third-party verification, and the model trails the absolute best from Anthropic and OpenAI. But if even half the claims hold, Kimi K3 forces a repricing conversation across the entire API market. That benefits every small operator running on thin margins. Test the endpoint, compare the bill, and decide for yourself.

Sources

- Constellation Research: Moonshot AI Launches Kimi K3 - FelloAI: Kimi K3 Overview - ToSea: Kimi K3 Complete Guide - 163.com: Kimi K3 Release Coverage - LM Market Cap: Kimi K3 Profile