Kimi K3 Explained: Moonshot's 2.8T Open Weight Flagship

On July 16, 2026, Moonshot AI released Kimi K3, and the headline writes itself: it is the largest model ever announced with an open weight release on the calendar. Roughly 2.8 trillion parameters, a 1 million token context window, launch day benchmarks that poke at the frontier labs, and a promise that the full weights land on Hugging Face by July 27.

Here is everything actually known so far, what is claim versus fact, and what a 2.8T model realistically means for people like us who prefer running AI on hardware we own.

What Kimi K3 Is

K3 is a new architecture Mixture of Experts (MoE) model. The 2.8 trillion number is the total parameter count; as with every MoE, only a fraction of those experts activate per token, which is what keeps inference costs sane. Moonshot positions it squarely at long horizon coding and agent workloads: big context, tool calling, structured outputs, and image input.

Two variants shipped at launch:

• K3 Max: the main chat and agent model.
• K3 Swarm Max: tuned for large scale parallel processing, think fleets of agents fanning out over a task.

The rollout started on Kimi Code and inside the Kimi app (with a ¥199 subscription tier as the entry point), plus the Moonshot open platform API. A leaked promo page on Moonshot's own site tipped the release a day early, which is why half of AI Twitter was already arguing about it before the official post went up.

The Spec Sheet

The Benchmark Claims (Read the Fine Print)

Moonshot's self reported numbers have K3 mostly beating Claude Opus 4.8 max and GPT 5.5 high, while losing to Claude Fable 5 and GPT 5.6. If that holds up under independent testing, an open weight model sits comfortably inside the frontier pack, a place open models have only brushed against before.

Two honest caveats. First, these are launch day, vendor picked benchmarks; independent verification is thin this early, and launch week numbers have a history of shrinking in the wash. Second, "mostly beating" is doing some work in that sentence: the wins are concentrated in coding and agentic tasks, which is exactly what Moonshot optimized for.

Still, the direction matters more than the decimals. The gap between open weight and closed frontier models keeps narrowing, and it is Chinese labs (Moonshot, DeepSeek, Qwen, Zhipu) doing most of the narrowing. For anyone who cares about running capable models outside of closed clouds, that trend is the story.

Pricing: Cheap for a Frontier Model, Not Cheap

The API costs $3 per million input tokens ($0.30 on a cache hit) and $15 per million output tokens, reasoning tokens included. That undercuts most Western flagships, but notice the drift: the K2 series launched dramatically cheaper. Frontier scale MoEs are expensive to serve even when the weights are open, and the era of Chinese models being nearly free at the API level is visibly ending.

Right now the only place running K3 is Moonshot itself. OpenRouter lists the model, but every request is forwarded to Moonshot's servers; there is no second provider, because nobody else has the weights yet. Which brings us to the interesting part.

The Open Weights Date Is the Real Event

Moonshot says the full K3 weights will be public by July 27, 2026. The K2 precedent suggests a Modified MIT license, which permits pretty much everything that matters, including hosting it commercially and fine tuning it.

Why that date matters more than launch day:

• Independent hosting. Once the weights are out, neutral inference providers can serve K3, prices fall, and your prompts no longer have to route to the model maker's own servers.
• Real benchmarks. The community can verify the launch claims instead of taking screenshots on faith.
• Derivatives. Quantizations, distills, fine tunes, and yes, eventually uncensored variants. The abliteration pipeline that produced today's best uncensored models needs weights to work on. No weights, no freedom.

Until the 27th, treat K3 as a very impressive hosted product with an open promise attached.

Can You Run It Locally?

Short version: no, and it is worth being straight about that. A 2.8T parameter model at 4 bit quantization is around 1.4 terabytes of weights before you allocate a single byte of KV cache for that 1M context. That is multi node datacenter hardware, not a gaming PC, and not even a maxed out Mac Studio.

We did the full math, including what the weights drop changes and what you can realistically run instead, in a companion piece: Can You Run Kimi K3 Locally? The one sentence preview: your local machine runs excellent 8B to 70B open models today, and K3 class quality is what hosted access is for.

If you are wondering how K3 compares to its very runnable smaller sibling, we also lined it up against K2.6: Kimi K3 vs Kimi K2.6.

What This Means for Local AI

Every time a frontier class model goes open weight, the local ecosystem inherits the benefits about six months later: the techniques, the distills, the tooling. K2's release pushed open agent models forward across the board. K3's architecture doing the same is the realistic upside for those of us running models at home.

And in the meantime, the practical setup does not change: run capable open models on your own hardware for privacy and zero refusals, and keep an eye on the July 27 weights drop. If you are new to local models, our beginner guide to running AI locally and the uncensored setup guide get you from zero to a working chat in minutes with Locally Uncensored.

FAQ

What is Kimi K3?

Moonshot AI's flagship model, released July 16, 2026: a Mixture of Experts model with roughly 2.8 trillion total parameters and a 1 million token context window, aimed at coding and agent workloads. It shipped as K3 Max (chat and agents) and K3 Swarm Max (parallel processing).

Is Kimi K3 open source?

Announced as open weight, but the weights are not out yet. Moonshot promises them by July 27, 2026, expected under a Modified MIT license like the K2 series.

How much does the Kimi K3 API cost?

$3 per million input tokens ($0.30 with a cache hit) and $15 per million output tokens, including reasoning tokens. OpenRouter charges the same because it routes to Moonshot.

Is Kimi K3 better than Claude or GPT?

Per Moonshot's own benchmarks it mostly beats Claude Opus 4.8 max and GPT 5.5 high and loses to Claude Fable 5 and GPT 5.6. Independent numbers are not in yet; treat it as a strong claim.

Can I run Kimi K3 on my own computer?

Realistically no. The weights alone are around 1.4 TB at 4 bit. See our full breakdown of the hardware math and the local alternatives worth running instead.

Getting Started with Local AI Today

K3 sized models need a datacenter, but everything up to about 70B runs beautifully at home, uncensored, private, and free:

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
# Windows: setup.bat | Linux: ./setup.sh

Or grab the installer from the releases page, open the Model Manager, and one click a model. When the K3 weights drop on the 27th, we will cover the derivatives worth caring about.

Locally Uncensored is AGPL-3.0 licensed and free to use. Built by PurpleDoubleD.