Moonshot’s Kimi K3 Nears the AI Frontier, Then Charges Frontier Prices

Moonshot AI released Kimi K3 on Thursday, a 2.8 trillion parameter model the Beijing startup calls the first open model ever to reach that scale. It lands with benchmark scores that crowd the American frontier, a 1 million token context window, and a catch: running it now costs five times what the company’s own prior model charged, and the actual downloadable weights are still more than a week away.

The launch has been read across the industry as proof China closed the gap with Anthropic and OpenAI. Less discussed is what closing that gap cost Moonshot: the rock bottom pricing and instant openness that made Chinese open models a competitive threat in the first place.

Moonshot Calls K3 the World’s First Open 3-Trillion-Class Model

Moonshot AI, the Beijing based startup backed by Alibaba and Tencent, described its new flagship plainly on its own blog: “Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window.”

The company calls it the world’s first open 3T-class model, built for long horizon coding, knowledge work and reasoning. It is compatible with the OpenAI SDK, which lowers the barrier for developers already building on ChatGPT or Claude toolchains.

Moonshot’s own timeline chart, cited by outlets that reviewed the launch materials, places K3 well clear of every other Chinese open model on the market:

• Kimi K3 – 2.8 trillion parameters, released July 16
• DeepSeek V4 Pro – roughly 1.6 trillion parameters, the previous scale leader
• Xiaomi’s largest model – about 1.02 trillion parameters
• Alibaba’s largest model – about 397 billion parameters

That makes K3 close to twice the size of DeepSeek’s biggest release and nearly three times the size of Moonshot’s own K2 generation, which held at 1 trillion total parameters for roughly a year.

The Architecture Behind the Leap

Moonshot did not just add parameters. It rebuilt how the model routes information, using two techniques it developed internally and had previously published as open research.

• 2.5 times the overall scaling efficiency of Kimi K2, according to Moonshot’s own platform documentation, meaning it converts compute into capability more effectively than its predecessor.
• 16 of 896 experts activate for any given token under a new framework called Stable LatentMoE, keeping activation sparsity under 2 percent.
• 6.3 times faster decoding in million token contexts, a gain the company attributes to Kimi Delta Attention, its hybrid linear attention mechanism.
• Roughly 25 percent higher training efficiency from Attention Residuals, which Moonshot describes as a drop-in replacement for standard residual connections.

Running the finished model is its own undertaking. Moonshot recommends a deployment spanning at least 64 accelerators, and its API quickstart documentation notes the team is still working with inference partners to align technical details before the model ships more broadly across the open source ecosystem.

Landing Days Before Beijing’s Biggest AI Stage

The timing is not incidental. K3 arrived just ahead of the 2026 World Artificial Intelligence Conference in Shanghai, where Chinese President Xi Jinping is expected to lay out Beijing’s AI priorities. Moonshot’s domestic rival DeepSeek is also expected to release an updated model in short order.

The release also marks a comeback of sorts. Moonshot’s market position had eroded over the prior eighteen months as DeepSeek’s rise pulled attention and developer mindshare away from the Kimi line. K3 is the company’s attempt to reclaim the spotlight, arriving alongside reports that Moonshot is raising fresh capital at a valuation near $31.5 billion, up from the $20 billion valuation it reached after raising $2 billion in May.

It fits a wider pattern of Chinese labs pushing agentic AI into places American companies have been slower to reach, including AI agents now handling shopping tasks once done by humans. Moonshot’s earlier Kimi models were already showing up inside Silicon Valley tools: coding startup Cursor used Kimi to help build its Composer 2 agent, and DoorDash has been routing lower level engineering work to Kimi K2.6, according to the company’s chief technology officer.

How Close Does Kimi K3 Really Get to Claude and GPT-5.6 Sol?

Independent testing puts Kimi K3 just behind the two leading proprietary models and ahead of everything else, including Anthropic’s previous flagship. It is not the top model on any major composite benchmark, but it is close enough to force a rethink of how far ahead closed source labs actually are.

On GDPval-AA v2, a benchmark from analytics firm Artificial Analysis measuring real world tasks across 44 occupations, the results line up like this:

Claude Opus 4.8, Anthropic’s prior flagship, scored 1,600 on the same GDPval-AA v2 test, meaning Kimi K3 now outperforms a model that sat at the frontier only weeks ago. On Artificial Analysis’s broader Intelligence Index, K3 scores just above 57, a few points behind Fable 5 and GPT-5.6 Sol but ahead of Opus 4.8 and Claude’s Sonnet tier. On Arena’s Frontend Code leaderboard, K3 jumped from eighteenth place under the K2.6 generation straight to first, ahead of Fable 5.

Kimi K3 is a “big moment, with multiple implications for the entire industry.”

Sriram Krishnan, a former senior White House policy adviser on AI, offered that assessment shortly after launch. Mozilla’s chief technology officer Raffi Krikorian struck a similar note, telling Axios that “right now, it’s a U.S. versus China question,” and arguing that American AI labs would have little reason to lobby against open weight models unless they saw them as genuine competition.

A Fivefold Price Jump Erases the Old Kimi Pitch

Here is where the story turns. Moonshot’s earlier models built their reputation partly on being cheap. K3 is not.

The new model costs $3 per million input tokens, $0.30 per million cached input tokens, and $15 per million output tokens, according to Moonshot’s own pricing documentation. That is roughly five times what Kimi K2.6 charged for the same volume of output just months ago.

That last row matters most. Anthropic is reportedly planning to raise Claude Opus 4.8’s price in September to $3 per million input tokens and $15 per million output tokens, the exact same numbers Moonshot just attached to K3. The model built to undercut Western pricing now sits precisely where Anthropic’s own repriced flagship is headed. Moonshot says its annual recurring revenue already exceeds $200 million, so the company is betting enterprise buyers will pay premium rates for a model it still calls open.

The Open Label Still Carries Asterisks

Weights, Still More Than a Week Out

Moonshot says full model weights will ship by July 27, a window that stretched past this article’s publication with nothing yet posted. As of July 17, Moonshot’s Hugging Face organization listed only its older K2 series checkpoints. Artificial Analysis has classified the currently accessible K3 service as proprietary because there is no weight file to independently inspect yet.

The company has also not disclosed how many of K3’s 2.8 trillion parameters activate per token, a figure that matters for anyone trying to estimate real hardware costs. Nor has it set a publication date for the technical report that would lay out training data and evaluation methods in full.

What We Know

• Kimi K3 is live now through Moonshot’s website, apps, Kimi Code and its API, priced at $3 input and $15 output per million tokens.
• Moonshot has committed to a July 27 date for releasing the full weights.
• Independent evaluators, including Artificial Analysis and Arena, have already run and published their own benchmark numbers.

What’s Unconfirmed

• The active parameter count per token has not been disclosed.
• No technical report or license terms have been published yet.
• Whether launch day benchmarks hold up once outside researchers can run the weights themselves.

The Anthropic Shadow

K3 also launches under a cloud that predates it by months. In February, Anthropic said it had identified industrial scale distillation attacks targeting Claude, accusing DeepSeek, Moonshot and MiniMax of generating over 16 million exchanges with Claude through roughly 24,000 fraudulent accounts, in violation of its terms of service and regional access restrictions.

Distillation itself is not inherently improper. It simply means training a smaller or newer model on a stronger rival’s outputs rather than building every capability from scratch, a practice Anthropic itself acknowledges is legitimate when a lab does it to its own models. What Anthropic alleged was systematic extraction from a competitor’s product through accounts created to evade its access rules. The Trump administration has since labeled the practice “adversarial” and pledged to crack down on it, and K3’s benchmark strength in exactly the coding and reasoning categories Anthropic flagged is likely to reheat that argument.

What Breaks After July 27

The real test starts once outside researchers can download the weights and run K3 on hardware nobody at Moonshot controls. Vendor benchmarks are self reported by definition, and several early testers already flagged slow reasoning traces and long runtimes on harder agent tasks even as they praised the model’s coding and design output.

There is a precedent for how fast this can move markets. In January 2025, DeepSeek’s R1 release matched a leading American model at a fraction of the price and wiped out roughly a trillion dollars in tech stock value within days. K3 is a bigger, pricier model than R1 was, but the underlying question, whether Washington’s export controls are actually slowing Chinese frontier AI, is the same one investors and policymakers will be asking again once the weights are public and DeepSeek’s own next release lands.

Frequently Asked Questions

What is Kimi K3?

Kimi K3 is Moonshot AI’s flagship large language model, built on 2.8 trillion parameters with a 1 million token context window and native image understanding. Developers can already select it inside Kimi Code using the model ID k3 and the /model command, ahead of the broader open weight release.

Is Kimi K3 actually open source?

Not yet in a downloadable sense. Moonshot calls it an open model and promises full weights by July 27, but as of this writing no checkpoint has appeared on Hugging Face and Artificial Analysis still classifies the hosted service as proprietary until that changes.

How much does Kimi K3 cost to use?

The API charges $3 per million input tokens, $0.30 per million cached input tokens and $15 per million output tokens. Web search calls carry an additional $0.015 charge per call, on top of standard token pricing.

How does Kimi K3 compare with Claude and GPT-5.6 Sol?

It trails both on Moonshot’s own admission and on independent scores, but not by much. On Arena’s Frontend Code leaderboard, K3 posted a 76 percent pairwise win rate against Claude Fable 5’s 63 percent and GPT-5.6 Sol’s 58 percent in head to head developer preference tests.

When do Kimi K3’s open weights actually ship?

Moonshot has committed to July 27, 2026. Infrastructure provider vLLM has already confirmed it added runtime support for Kimi Delta Attention’s prefix caching ahead of that date, suggesting the ecosystem is preparing for the release even before the files land.

Who is behind Moonshot AI?

Moonshot AI is a Beijing based startup backed by Alibaba and Tencent. It raised $2 billion in May at a $20 billion valuation and is reportedly now raising capital again at close to $31.5 billion, with annual recurring revenue that a company financial adviser says has already passed $200 million.