Kimi K3 arrived on July 16, 2026, and it is the biggest bet Moonshot AI has ever made. The new flagship is a Mixture-of-Experts model with a reported 2.8 trillion parameters and a 1-million-token context window, and Moonshot is positioning it as an open-weight rival to Claude Opus 4.8 at a fraction of the price. The tech press summed up the moment in one line, China’s Moonshot is challenging Anthropic with a bigger, cheaper model.
There is a catch worth knowing before you get excited. The rollout is happening in real time, official documentation is live while independent benchmarks are still pending, so a lot of the numbers flying around are reported figures, not verified ones. This guide separates what Moonshot has confirmed from what testers are claiming, walks through the specs, pricing, and access, and shows how Kimi K3 stacks up against the models you already use.
The Key Takeaways
- Launched July 16, 2026 by Moonshot AI, live now through the Kimi app, Playground, and API.
- A Mixture-of-Experts model with a reported 2.8 trillion parameters and a 1M-token context window.
- Accepts text, image, and video input, with thinking always on and a tunable reasoning_effort control.
- Reported API pricing is $3 per million input tokens and $15 per million output tokens, undercutting most Western flagships.
- Positioned as an open-weight Claude Opus 4.8 rival, though independent benchmarks and the public weights download are still pending.
What is Kimi K3?
Kimi K3 is Moonshot AI’s flagship large language model, launched on July 16, 2026. It uses a Mixture-of-Experts design with a reported 2.8 trillion total parameters and a 1-million-token context window, and it accepts text, image, and video input. Moonshot is releasing it as an open-weight model, which means developers will be able to download and modify it, a sharp contrast to the closed flagships from OpenAI and Anthropic.
If you saw conflicting reports about whether K3 was actually out, you were not imagining it. Early on July 16 some outlets still described the model as shipping “in the coming days,” while Moonshot’s own platform already carried a live K3 quickstart with Playground and API access. The short version, the model went live on the evening of July 16, and the confusion was just the usual gap between a staggered rollout and the press catching up.
| Spec | Kimi K3 |
|---|---|
| Developer | Moonshot AI |
| Released | July 16, 2026 |
| Architecture | Mixture-of-Experts (MoE) |
| Parameters | ~2.8 trillion (reported) |
| Context window | 1,000,000 tokens |
| Input types | Text, image, video |
| Reasoning control | reasoning_effort (currently max only) |
| License | Open weights (expected, not yet published) |
Kimi K3 specs and architecture
The headline number is scale. At a reported 2.8 trillion parameters, Kimi K3 is described as the largest AI model China has produced, more than double the roughly 1 trillion parameters in the previous Kimi K2.6. Because it is a Mixture-of-Experts model, only a slice of those parameters activates on any single request, which is how Moonshot keeps inference costs low enough to price the model aggressively.
A 1-million-token context window
Kimi K3 handles up to 1 million tokens in a single context, roughly four times the 256K window of K2.6. That is enough to hold an entire codebase, a stack of long documents, or a lengthy multi-step agent run without losing the thread. Long context is the feature Moonshot is leaning on hardest, and it lines the model up directly against the largest windows offered by rivals.
Multimodal input and always-on reasoning
K3 accepts text, images, and video, which pushes it past the text-first K2 line into proper multimodal territory. Reasoning is always on, and Moonshot exposes a reasoning_effort parameter to tune how hard the model thinks. At launch that control only supports the max level, with more levels promised soon, and reasoning tokens are billed as output. There is also built-in web search, charged separately per call.
How it evolved from Kimi K2
K3 is the next step after a busy K2 family that ran through K2, K2.5, K2.6, and the coding-focused K2.7. Each of those was an open-weight release, and K2.6 in particular earned a reputation as one of the strongest open coding models available. If you want the full lineage, our breakdown of Kimi K2.7 Code and how the K2 line handles coding is the place to start. K3 more than doubles the parameter count and quadruples the context on top of that foundation.
Kimi K3 benchmarks: what is verified and what is not
Here is where you should slow down. As of launch, Kimi K3 has no independent benchmark coverage. The numbers circulating come from Moonshot itself or from individual testers, so treat them as claims rather than confirmed results. Benchmark trackers that follow the model note they do not yet have enough non-generated data to give it a leaderboard rank, and Moonshot’s own Kimi K3 API quickstart lists specs without publishing benchmark tables.
The strongest first-party figure is on agentic web browsing. Moonshot reports Kimi K3 hitting 91.2% on BrowseComp in a single-agent setup without context compression or extra context-management tricks, a very high score if it holds up. On coding, community tester reads suggest K3 lands around an “Opus 4.7-plus” level and beats GPT-5.5 on some evaluations, while noting that GPT-5.6 Sol and Fable 5 still lead on the toughest terminal benchmarks. None of that is independently verified yet.
The honest takeaway, Kimi K3 looks like a genuine frontier-class contender on paper, but the “beats Opus 4.8” headlines are running ahead of the evidence. Give the independent labs a week and the picture will firm up. Until then, if you need a verified leader for repository-level coding today, Claude Opus 4.8 is still the safe pick.
Kimi K3 vs Claude Opus 4.8 and GPT-5.5
The comparison everyone is running is Kimi K3 vs Claude Opus 4.8, and on paper it is a study in opposites. K3 is enormous and open; Opus 4.8 is comparatively compact, closed, and battle-tested. Anthropic has never published Opus 4.8’s parameter count, but on-chain and community estimates put it far below K3’s scale, which is exactly the point Moonshot is making about efficiency versus raw size.
| Model | Developer | Params (est.) | Context | Open weights |
|---|---|---|---|---|
| Kimi K3 | Moonshot AI | ~2.8 trillion (reported) | 1M tokens | Expected, not yet live |
| Claude Opus 4.8 | Anthropique | ~150–200 billion (unofficial) | 200K tokens | Non |
| GPT-5.5 | OpenAI | Not disclosed | Not disclosed | Non |
| Kimi K2.6 | Moonshot AI | ~1 trillion | 256K tokens | Yes |
What the table does not capture is track record. Opus 4.8 leads verified SWE-bench Pro results for repository-level engineering, and GPT-5.5 has the edge on terminal-heavy workflows. Kimi K3’s advantages are its open weights, 1M context, and price, which matter enormously if you are self-hosting or running huge context jobs. If you are choosing between flagships for a specific job, our guide on when to use which AI model breaks the decision down task by task.
Kimi K3 pricing
Reported Kimi K3 pricing is $3 per million input tokens and $15 per million output tokens, with built-in web search billed at $0.015 per call. Moonshot’s official page confirms it uses per-token billing with separate cache-hit and cache-miss input rates, but has not published the exact figures, so the $3 and $15 numbers come from third-party trackers and should be treated as reported. Even so, that undercuts most Western flagship models while offering a far larger context window.
| Model | Input / 1M | Output / 1M | Notes |
|---|---|---|---|
| Kimi K3 | $3.00 | $15.00 | Reported; web search $0.015/call |
| Kimi K2.7 Code | $0.95 | $4.00 | Coding-focused |
| Kimi K2.6 | $0.95 | $4.00 | Open weights |
| Kimi K2.5 | $0.60 | $3.00 | Open weights |
Note that K3 is a clear step up in cost from the K2 line, which reflects its bigger footprint and always-on reasoning. If your workload does not need frontier-level scale, K2.6 or K2.7 Code remain much cheaper and are already fully open. There is no confirmed free tier for K3 itself yet, though the Kimi app offers free access to the wider Kimi model family with usage limits.
How to use Kimi K3
There are three ways to reach Kimi K3, and which one fits depends on whether you want a chat window, an API, or the raw weights. Here is the practical order.
- Kimi app and Playground. The fastest route is the Kimi app or the web Playground, where K3 is selectable now for logged-in users. This is the no-code option for testing the model on your own prompts.
- The API. Developers can call K3 through Moonshot’s Kimi API platform, with a quickstart already live. It supports function calling, JSON-schema structured output, and the reasoning_effort control. K3 is also appearing on aggregators like OpenRouter for teams that route across providers.
- Open weights. Moonshot has released past Kimi models on Hugging Face, and K3 open weights are widely expected. As of launch there is no public model card, license, or download yet, so self-hosting is the “soon” option, not the “today” one.
If you would rather not juggle a separate subscription for every new model, apps like Fello let you use powerful current AI models in one place, which is handy while a launch like this one settles. For a wider view of the open field, our roundup of the best open-source AI models available right now puts K3’s lineage in context, and if you are weighing Chinese open models specifically, our explainer on what GLM is and how it competes is a useful companion.
Why Kimi K3 matters
Kimi K3 is not just another model drop. If the open weights ship as expected, it would be the largest open-weight model from China to date, handing developers frontier-scale capability they can download, inspect, and run themselves. That is a direct challenge to the closed, premium-priced flagships that have defined the top of the market, and it lands while the momentum behind capable open models is already building.
The business context makes the ambition clear. Moonshot’s Kimi family reportedly passed $300 million in annualized recurring revenue by mid-June, and, according to TechCrunch’s launch report, the company is raising fresh capital at a reported $31.5 billion valuation, up sharply from the $20 billion figure attached to its round earlier in 2026. A well-funded lab shipping a giant open model at aggressive prices is exactly the kind of pressure that reshapes what everyone else charges. To see how that reshapes model choice, our guide on getting the most out of GPT-5.6 shows how fast the frontier is moving on the closed side too.
The bottom line
Kimi K3 is a big moment, a 2.8-trillion-parameter, 1M-context, open-weight model priced to undercut the West and aimed squarely at Claude Opus 4.8. What it is not, yet, is a proven benchmark champion. The specs and price are real and confirmed enough to act on; the “beats everything” claims need independent testing before you rebuild your stack around them.
Our recommendation, try K3 in the Playground or via the API today to feel out its long-context and reasoning strengths, but keep a verified model like Opus 4.8 in the loop for mission-critical coding until the independent numbers land. Check back as the open weights and third-party benchmarks arrive, that is when Kimi K3’s real place in the pecking order will become clear.
FAQ
Is Kimi K3 out yet?
Yes. Moonshot AI launched Kimi K3 on July 16, 2026, and it is live through the Kimi app, the web Playground, and the API. Some early reports still called it upcoming, but the model went live the same evening.
How many parameters does Kimi K3 have?
Kimi K3 has a reported 2.8 trillion total parameters in a Mixture-of-Experts design, making it the largest AI model China has produced. Only a fraction of those parameters activate per request, which keeps inference costs down.
Is Kimi K3 better than Claude Opus 4.8?
On paper Kimi K3 is far larger and open-weight, and Moonshot claims parity or better on some evals. But independent benchmarks are still pending, so Claude Opus 4.8 remains the verified leader on repository-level coding for now.
How much does Kimi K3 cost?
Reported API pricing is 3 dollars per million input tokens and 15 dollars per million output tokens, with web search billed at 0.015 dollars per call. Moonshot confirms per-token billing but has not published official figures, so treat these as reported.
Is Kimi K3 open source?
Moonshot is releasing Kimi K3 as an open-weight model, following the K2 line on Hugging Face. As of launch there is no public model card, license, or weights download yet, so self-hosting is expected soon rather than available today.
