Moonshot’s Kimi K3 Steps Into the Gap the US Just Opened

Moonshot AI released Kimi K3 on July 16, a 2.8 trillion-parameter open-weight model, one month after Washington forced Anthropic to pull its most capable AI system from foreign users worldwide. The Beijing startup calls it the world’s first open 3T-class model, built for long-horizon coding, deep reasoning and a 1 million-token context window that lets it hold far more information in a single prompt than earlier generations.

The timing does more than look awkward for U.S. policymakers. Kimi K3 needs no nationality check, no export license and no verification system to download. That is precisely the gap Washington’s own order tore open when it barred foreign nationals, including Anthropic’s own noncitizen employees, from touching the company’s newest model.

Moonshot Ships the Biggest Open Model Anyone Has Seen

Kimi K3 is a sparse mixture-of-experts system that activates just 16 of its 896 experts on any given token, roughly 1.8% of the total pool, which is why a 2.8 trillion-parameter model can run at all without dense-model inference costs. Moonshot says two internal innovations, Kimi Delta Attention and Attention Residuals, deliver close to a 2.5 times gain in scaling efficiency over its predecessor, Kimi K2.6.

Independent evaluators have weighed in even before the full weights ship. Arena.ai ranked Kimi K3 first for web interface building. Vals AI placed it second overall, behind Anthropic’s Fable 5 but ahead of OpenAI’s GPT-5.6 Sol. Artificial Analysis found its performance in the same range as GPT-5.5 and Claude Opus 4.8 on complex, multi-step reasoning work, and recorded an Elo jump of 732 points over Kimi K2.6 on its private long-horizon knowledge evaluation, the largest single-generation leap it has logged for the Kimi line.

None of that is verifiable yet in the strictest sense. Moonshot has only released K3 through its website and API; full weights are due July 27. Every published benchmark until then is a number the company or a partner reported, not one an outside lab has reproduced against the actual model file.

The Ban That Built K3’s Market

Understanding why K3 lands the way it does requires going back to June 12, when the U.S. Commerce Department ordered Anthropic to suspend all access to Claude Fable 5 and Mythos 5 for any foreign national, anywhere, including foreign nationals working inside Anthropic itself. The directive arrived just three days after Fable 5’s public debut and cited a jailbreak technique the government said exposed Mythos’s underlying cybersecurity capabilities.

Anthropic complied but pushed back hard on the reasoning.

We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.

Anthropic wrote that in its official statement on the suspension, adding that applying the same standard industry-wide would halt new model deployments across every frontier lab. The company said the jailbreak it reviewed was narrow and non-universal, and that the same technique could pull similar capability out of other publicly available models, including OpenAI’s GPT-5.5, which faced no such restriction.

Commentators outside Anthropic went further at the time. AI critic Gary Marcus argued the order made little sense given the government’s own stated goal of staying ahead of China, and warned it could push Chinese-born researchers at U.S. labs back toward China while making investors question the stability of American AI policy. Five weeks later, a Chinese lab shipped an open, downloadable alternative that anyone shut out by that order can simply run.

• Foreign nationals outside the U.S. – blocked from Fable 5 and Mythos 5 entirely under the directive.
• Foreign nationals inside the U.S. – including Anthropic’s own noncitizen staff, covered by the same suspension.
• Every other Claude model – Opus, Sonnet and Haiku remained available throughout, only the two newest systems were pulled.

Fortune’s coverage of Kimi K3’s launch noted the release could reopen debate over U.S. AI policy, arguing it might either loosen export controls to help American firms compete or embolden hawks pushing to restrict China’s AI sector further. Either way, K3 does not need Washington’s cooperation to reach the users the June order cut off.

Who Can Actually Run 2.8 Trillion Parameters

Open-weight is not the same as accessible. Ryan Fedasiuk, a fellow at the American Enterprise Institute, estimated that running a 2.8 trillion-parameter model locally would require computing infrastructure costing hundreds of thousands of dollars, putting real self-hosting out of reach for all but well-funded enterprises and governments. Moonshot itself recommends serving K3 on supernodes of 64 or more accelerators kept inside one high-bandwidth domain.

That is the irony sitting underneath Moonshot’s pitch: a model marketed as open to anyone is, in practice, open only to whoever can afford a data center. What it does open up is price competition against the same closed labs the export order was meant to protect.

Lian Jye Su, chief analyst at Omdia, told Reuters that enterprise interest in Chinese models comes down to cost as much as capability. “They can be run at a fraction of the cost that OpenAI charges its clients,” he said, while cautioning that raw parameter count is not a reliable proxy for model quality. K3’s own pricing bears that out unevenly: at $3 per million input tokens and $15 per million output tokens, it is the most expensive model any Chinese lab has released, yet still a third of what Anthropic charges for Fable 5 output.

How the Open-Weight Field Stacks Up

Bank of America analysts led by Alex Liu wrote in a note cited by CNBC that large-scale pre-training combined with architectural work can still produce step-change gains for Chinese flagship models despite ongoing compute constraints. That view lines up with what Moonshot’s own benchmarking shows against the GPU-limited hardware it discloses, including an Nvidia L20, the cut-down Ada-based accelerator sold into China under U.S. export rules. xAI has taken a narrower version of the same open approach, releasing Grok’s code while keeping a kill switch intact, a contrast to Moonshot’s plan to hand over full weights outright.

Anthropic’s Distillation Complaint Still Hangs Over Moonshot

K3 is not launching into a clean rivalry. Anthropic accused Moonshot in February of using 3.4 million Claude exchanges to train its own models through distillation, and K3 now benchmarks within a few points of the systems named in that complaint. Moonshot has not addressed the accusation publicly since the K3 launch.

The dispute sits alongside a wider compute story. China’s chip access has tightened steadily under U.S. export controls, even as the country’s exports of chips and AI-related hardware keep climbing; China’s exports jumped 27% in June on AI chip demand, underscoring how much industrial momentum is riding on exactly the hardware category Washington is trying to restrict.

What Changes When Weights Land on July 27

What we know:

• Kimi K3 launched July 16 with 2.8 trillion total parameters, a 1 million-token context window, and self-reported benchmarks placing it near Fable 5 and ahead of GPT-5.6 Sol on select tasks.
• Full weights, not just API access, are scheduled for release on July 27.
• Pricing at $3 input and $15 output per million tokens makes it the most expensive Chinese open model to date, still cheaper than Fable 5.

What’s unconfirmed:

• Every benchmark number remains Moonshot-reported or drawn from limited API testing; no outside lab has reproduced results against the actual model file.
• Exact licensing terms for the July 27 weight release have not been published; Moonshot’s prior K2 model shipped under a Modified MIT license.
• Whether foreign-national access to Fable 5 and Mythos 5 has been restored since June remains unclear; Anthropic said only that it was working to resolve the suspension.

Once the weights are public, independent labs can finally test Moonshot’s claims against the model itself rather than an API response. That is also the moment enterprises priced out of self-hosting will decide whether to rent capacity from third-party inference providers instead of building their own supernodes, a decision K3’s own cost structure was clearly built around.

Frequently Asked Questions

What does open-weight mean for Kimi K3?

Open-weight means the trained model file itself can be downloaded, run and modified, unlike a closed API-only system such as Fable 5. It does not necessarily mean the training data or full source code are public. Moonshot released its prior Kimi K2 model under a Modified MIT license, and the exact license for K3’s July 27 weight drop has not yet been published.

Can smaller developers access Kimi K3 without buying their own hardware?

Yes. Beyond Moonshot’s direct API, third-party inference marketplaces such as OpenRouter already list Kimi K3, letting developers query the model without signing up for a Moonshot account or owning any of the underlying infrastructure.

Are Anthropic’s Fable 5 and Mythos 5 back online yet?

There is no confirmed update on restoration. Anthropic’s June statement said only that it was working to restore access “as soon as possible” after the export-control suspension. All other Claude models, including Opus 4.8, remained available throughout and were never part of the order.

How does Kimi K3’s price compare to GPT-5.6 Sol on a per-task basis?

Artificial Analysis found Kimi K3’s cost per task runs about $0.94, close to GPT-5.6 Sol’s $1.04 and roughly half of Claude Opus 4.8’s $1.80, though still higher than most open-weight peers on a per-task basis.