Kimi K3 is the newest AI model from Chinese startup Moonshot AI, released on July 16, 2026. At roughly 2.8 trillion parameters it's the largest open-weight model ever released, with a 1M-token context window, native multimodal input, and reasoning on by default. On the overall Artificial Analysis Intelligence Index it lands around 4th (behind Claude Fable 5 and GPT-5.6 Sol), but it takes first place on the Frontend Code Arena and is by far the cheapest of the frontier models at $3 in / $15 out per million tokens. You can try it free in the Kimi app on the $0 Adagio plan.
If you've been asking what is Kimi K3, here's the short version: it's a frontier-class AI model from Moonshot AI that Chinese and US press covered heavily in mid-July 2026 because it closes much of the gap with the best models from OpenAI and Anthropic while being open-weight and dramatically cheaper to run. This guide walks through what K3 actually is, how it benchmarks against Claude and GPT, what it costs, and how to try it today, all current as of July 2026.
Kimi K3 at a glance
| Attribute | Kimi K3 |
|---|---|
| Maker | Moonshot AI (China) |
| Released | July 16, 2026 |
| Parameters | ~2.8 trillion (open-weight, Mixture-of-Experts) |
| Context window | 1M tokens |
| Input types | Text and images (multimodal) |
| Reasoning | On by default on every query |
| API price | $3 / 1M input · $15 / 1M output |
| Free access | Yes, Kimi app "Adagio" plan ($0) |
| Open weights | Full weights promised by July 27, 2026 |
Specs, prices, and benchmark numbers reflect vendor and press reporting as of July 2026 and change fast. Check Moonshot's own model and pricing pages before relying on exact figures.
What Kimi K3 actually is
Kimi K3 is the fifth major Kimi release in under a year, following the K2 line and the June 2026 coding-focused K2.7-Code. Where it stands apart is scale: at about 2.8 trillion parameters it's the biggest open-weight large language model anyone has shipped. "Open-weight" means Moonshot publishes the model's trained parameters so developers can download, self-host, and fine-tune it, rather than only renting it through an API. Moonshot said it would release the full weights and a technical report by July 27, 2026.
Under the hood it's a Mixture-of-Experts (MoE) design, so even though the total parameter count is enormous, only a fraction of the network activates for any given token. That's what keeps a model this large usable at a sane cost. K3 also ships with a 1M-token context window, native multimodal input (text and images), and a reasoning mode that runs by default on every query rather than being a separate "thinking" toggle.
How Kimi K3 compares to Claude and GPT
The headline framing in the press was "China's biggest open model closes the gap with US rivals," and the benchmarks broadly support a nuanced version of that.
On the composite Artificial Analysis Intelligence Index, K3 scores about 57, which puts it in the top tier: essentially level with Claude Opus 4.8 (~56) and just behind GPT-5.6 Sol (~59) and Claude Fable 5 (~60). Reported placements vary by source and shift as new models launch, so it's frontier-adjacent rather than the outright leader on general intelligence.
Where it genuinely leads is coding and front-end work. K3 debuted at number one on Arena's Frontend Code Arena, passing Claude Fable 5 and GPT-5.6 Sol in blind developer testing. It also posted the strongest open-weight GPQA Diamond result to date at 93.5%.
Read benchmarks skeptically
Composite indexes and arenas measure different things, and vendor-friendly framing is common on launch day. A model can top a front-end coding arena and still trail on long-context reasoning or tool use. If a specific workload matters to you, test K3 against your current model on your own tasks before switching.
Here's the practical comparison most people care about:
| Model | Intelligence Index | API price (in / out per 1M) | Open weights |
|---|---|---|---|
| Kimi K3 | ~57 | $3 / $15 | Yes |
| Claude Fable 5 | ~60 | Higher (closed) | No |
| GPT-5.6 Sol | ~59 | Higher (closed) | No |
| Claude Opus 4.8 | ~56 | Higher (closed) | No |
The story the table tells: K3 trades a few points of top-line capability for being far cheaper than the closed US flagships (independent estimates put its per-task cost at roughly half that of Claude Opus 4.8), while also being the only one you can download and run yourself. For a fuller picture of the closed US flagships, see our explainers on GPT-5.6 Sol and how the big chatbots stack up in ChatGPT vs Claude vs Gemini.
The pricing catch worth knowing
The $3 / $15 sticker price is the cheapest of the frontier models, but there's a wrinkle. K3 is verbose: it tends to generate more tokens to reach an answer than its predecessors, and its per-token price is notably higher than the older, cheaper K2.6. In practice, per finished task it still comes out cheaper than the top closed models, but "cheap per token" and "cheap per task" aren't automatically the same thing. If you're budgeting API spend, measure cost on completed tasks, not raw token rates.
For most non-developers, though, pricing is simpler. The Kimi app is free on the $0 "Adagio" plan, which includes basic chat, file uploads, and web search. Paid Moonshot plans run from about $19/month up to $199/month depending on how much agentic and coding work you need, with roughly 20% off if you pay annually.
How to try Kimi K3 free
You don't need to be a developer to use it. The fastest path is the consumer app:
- Use the app or website. Sign in to the Kimi app and select K3 as the model. The free Adagio tier is enough for chat, uploading a document, and web-connected answers.
- Feed it long inputs. The 1M-token context is the standout feature for everyday use: you can drop in a long PDF, a codebase, or a pile of notes and ask questions across the whole thing.
- For coding, use the API or CLI. Developers can call K3 through the Kimi API or run it via Moonshot's Kimi Code CLI, and self-hosters can wait for the open weights (due July 27, 2026) to run it on their own hardware.
If you're comparing AI coding assistants specifically, our roundup of Cursor vs Copilot vs Claude Code covers where an open model like K3 fits against the popular editor-based tools, and our list of the best free AI tools in 2026 puts K3's free tier in context with everything else worth trying.
Should you switch to Kimi K3?
For budget-sensitive, code-heavy, or self-hosting use cases, K3 is genuinely compelling: front-end coding leadership, a huge context window, and the lowest frontier price, all in a model you can run yourself. For general assistant work where you want the highest possible reasoning quality and you're not counting tokens, Claude Fable 5 and GPT-5.6 Sol still edge ahead on the composite index. And because K3 is verbose and brand-new, it's worth a real side-by-side on your own tasks before you commit a workflow to it.
The bigger takeaway is directional: an open-weight model this close to the US frontier, at a fraction of the cost, changes the math for anyone building on top of AI.
Frequently Asked Questions
Is Kimi K3 free to use?
Yes, in the consumer Kimi app. The $0 Adagio plan includes basic chat, file uploads, and web search using K3. Heavier agentic and coding usage requires a paid Moonshot plan (roughly $19–$199/month), and developers pay per token through the API.
Who makes Kimi K3?
Moonshot AI, a Beijing-based AI startup. K3 is its most capable model to date and was released on July 16, 2026.
Is Kimi K3 better than Claude or ChatGPT?
It depends on the task. On the overall Artificial Analysis Intelligence Index, K3 (~57) sits just behind Claude Fable 5 (~60) and GPT-5.6 Sol (~59). But it leads on the Frontend Code Arena and costs far less, so for coding and cost-sensitive work it can be the better pick.
What does "open-weight" mean for Kimi K3?
It means Moonshot publishes the model's trained parameters so anyone can download, self-host, and fine-tune it. Full K3 weights were promised by July 27, 2026. That's different from closed models like Claude and GPT, which you can only access through their APIs.
How big is Kimi K3's context window?
1 million tokens, which is large enough to hold long documents, entire codebases, or extensive chat history in a single prompt.
Conclusion
Kimi K3 is the clearest sign yet that open-weight AI is closing in on the closed US frontier: near-top-tier intelligence, best-in-class front-end coding, a 1M-token context window, and the lowest price of any frontier model, all in something you can eventually run yourself. It won't unseat Claude Fable 5 or GPT-5.6 Sol at the very top of general reasoning, but for coding, long-context work, and tight budgets it's a serious option. The easiest next step is to open the Kimi app, switch to K3 on the free plan, and try it on a task you'd normally hand to your current model, then compare the results for yourself.