What Is Kimi K3? Moonshot AI's New Model Explained

An open model just crashed the frontier club.

On July 16, 2026, Beijing-based Moonshot AI released Kimi K3 — a 2.8-trillion-parameter model that immediately ranked third on Artificial Analysis, behind only Claude Fable 5 and GPT-5.6 Sol.

The Hacker News release thread hit 2,086 points and 1,204 comments in a day. Three days later, Moonshot had to suspend new subscriptions because demand outran their GPUs.

That almost never happens. Here's what Kimi K3 actually is, what it's good at, and how to try it for free.

Kimi K3 in one paragraph

Kimi K3 is Moonshot AI's new flagship large language model. It's a Mixture-of-Experts system with 2.8 trillion total parameters — only 16 of its 896 experts activate per token — a 1-million-token context window, native vision input, and reasoning that is always on.

The headline claim: frontier-level intelligence at roughly Sonnet-tier prices, with the full weights promised as an open release on July 27, 2026. If that lands, it becomes the largest open-weight model ever published.

The specs that matter

Kimi K3 spec card: 2.8T parameters, 1M context, always-on reasoning, $3/$15 per 1M tokens

Spec	Kimi K3
Released	July 16, 2026
Architecture	MoE, 16 of 896 experts active per token (KDA + AttnRes attention)
Total parameters	2.8 trillion
Context window	1,048,576 tokens (1M), flat pricing at any length
Input	Text + images (native vision)
Reasoning	Always on (`reasoning_effort` defaults to max)
API price	$3 / 1M input tokens ($0.30 on cache hit), $15 / 1M output
Open weights	Promised by July 27, 2026

Two things stand out.

First, the pricing is flat. The millionth context token costs the same as the first — unusual for a 1M-context model, and aggressive on purpose.

Second, the reasoning can't be switched off. K3 thinks before every answer, which is part of why its outputs are strong and also why it's not fast. More on that below.

What the benchmarks say

Moonshot didn't position K3 as "a great open model." They positioned it as a frontier model that happens to be open. The numbers mostly back that up:

#3 overall on the Artificial Analysis intelligence index — only Claude Fable 5 and GPT-5.6 Sol score higher
#1 on Front-end Code Arena at 1,679 points, ahead of Fable 5 — a human-preference arena, not a synthetic benchmark
77.8% on Program Bench, edging out GPT-5.6 Sol (77.6%)
81.2% on Frontier SWE and the top spot on SWE Marathon, the long-endurance coding test

The community verdict matches. The most-upvoted take on Hacker News: "I've been running Kimi K3 alongside Claude on my normal coding work, and for all practical purposes I can't tell them apart."

Where K3 visibly shines is front-end, 3D, and creative coding. Reviewers keep one-shotting things that used to take a frontier closed model: a full macOS clone in the browser, an FPS shooter with Three.js graphics, WebGL galaxy simulations. One tester's armory-scene comparison against Claude Opus 4.8 wasn't close — K3 built a textured, lit 3D environment while Opus produced a sparse room.

The honest downsides

The Kimi K3 trade-off: smart and cheap but slow, versus fast but expensive

No model launch survives contact with real workloads unscathed. Two complaints keep recurring.

It's slow. At launch, Moonshot serves K3 at roughly 26 tokens per second. Claude Fable 5 runs around 41. Combined with always-on reasoning, K3 regularly takes 2-3x longer than a closed frontier model on the same task.

It overthinks. Several developers note that K3 spends heavy reasoning effort even on trivial tasks, which inflates both latency and output-token costs. Moonshot says lower reasoning-effort settings are coming.

And one asterisk: until the weights actually appear on Hugging Face on July 27, "open" is a promise, not a fact.

Why this release matters

The economics are the story. K3 charges $3 per million input tokens and $15 per million output. Claude Fable 5, the model it benchmarks closest to, charges $10 and $50.

That's not a discount — that's a different market. It explains why several billion-dollar teams are reportedly evaluating a wholesale move, and why Moonshot ran out of serving capacity in 72 hours.

For creators, the practical takeaway is simpler: the cost of building things with AI just dropped again. The indie developers using K3 to one-shot browser games still need voiceovers, narration, and sound design for what they build — that's the part a code model can't generate, and where an AI Audio Generator picks up.

How to try Kimi K3

The free path: go to kimi.com, pick K3 in the model selector, and use it in the chat app. That's the same model, no card required.

The developer path: the API is OpenAI-SDK compatible at api.moonshot.ai/v1 with model ID kimi-k3. If you already have an OpenAI or Anthropic toolchain, switching is a base-URL and key change. There's also a swarm mode that fans a task out to parallel sub-agents for long multi-step jobs.

A concrete weekend test: have K3 one-shot a small narrated demo or game prototype, then run the script through an AI Text to Speech tool and pick a narrator from an AI Voice Library to make it shippable. Total model cost will land in cents.

FAQ

Is Kimi K3 really open source? Open weights, promised for July 27, 2026. The weights weren't public at launch, and the license hasn't been named yet — worth watching before you build on the assumption.

Is Kimi K3 free to use? Yes, via the Kimi app and kimi.com. The API is pay-as-you-go at $3/$15 per million tokens.

Is Kimi K3 better than ChatGPT or Claude? On several coding and front-end benchmarks it's ahead; on overall intelligence indexes it sits just below Claude Fable 5 and GPT-5.6 Sol. It's meaningfully slower than both.

Who is Moonshot AI? A Beijing-based AI lab backed by Alibaba, known for the Kimi model family (K2.5, K2.6, K2.7 Code, now K3).

The takeaway: for the first time, "frontier model" and "open weights" describe the same release — and it costs a third of the closed alternative.