Moonshot AI has unveiled its biggest model yet, and the numbers are designed to turn heads. The Kimi K3.1 Pro launch, announced in the early morning of October 24, brings a 2.8-trillion-parameter architecture, a 1-million-token context window and a coding benchmark score that the company says beats GPT-6. For a Chinese lab that many Western developers still treat as an outsider, it is a deliberate bid for the top tier.
The model is the flagship of the Kimi family, and Moonshot is rolling it out in phases through November 4. At launch it is available only in the Kimi app and through Moonshot’s API, with a Highspeed edition promising the same capability at two to three times the output speed. The question now is whether developers outside China will actually switch, or whether K3.1 Pro stays a headline rather than a habit.
What Kimi K3.1 Pro actually is
Start with the size. Kimi K3.1 Pro is built on 2.8 trillion total parameters, with 104 billion parameters activated on each inference pass. That mixture-of-experts design is how modern labs keep giant models affordable to run: only a slice of the network fires for any given request, while the full model holds the knowledge.
Then there is the context window. At 1 million tokens, K3.1 Pro can swallow an entire large codebase, a stack of legal documents or a whole research archive in a single prompt. That matters for the agentic coding work Moonshot is targeting, where an assistant needs to hold a project in its head while it plans, edits and tests across dozens of files.
Moonshot also says it purged training data associated with cyberattacks before release. That detail reads as a direct answer to critics who worry about powerful models being misused, and it comes weeks after earlier Kimi models drew scrutiny over safety questions.
The benchmark claim that matters
On Terminal-Bench 4.0, a test of command-line coding skill, Moonshot reports a score of 65.7 for Kimi K3.1 Pro. That puts it above the 58.2 points the company cites for GPT-6 Astra, and just behind the 66.4 points cited for Claude Opus 5.5. In other words, Moonshot is claiming a seat at the same table as the two most talked-about Western models.
Two cautions apply. First, these are company-reported figures, not independent results, and vendors always pick the tests that flatter them. Second, a single coding benchmark does not capture writing quality, reasoning, safety or cost β the things that decide whether a model survives contact with real work. Still, the gap to GPT-6 on this test is large enough that developers will want to try it themselves rather than dismiss it.
The phased rollout through November 4 gives Moonshot time to watch for failures before everyone piles in at once. It also stretches the news cycle, keeping Kimi in conversation for two full weeks.
Why China, and why now
K3.1 Pro arrives in a crowded week for Chinese AI. MiniMax released the weights for its M3.1 Flash coding model on October 23, also with a 1-million-token window, at $0.15 per million input tokens and $0.60 per million output tokens. Domestic chipmakers Hygon and Cambricon both posted strong third-quarter growth. And on October 16, China added 23 categories of AI technology to a draft export control list, though open-source releases and API services fall outside that framework.
Read together, the pattern is clear: Chinese labs are competing on openness and price while Western labs compete on scale and brand. Moonshot’s bet is that a frontier-class coding model, available through an API, will pull in developers who are tired of rising prices elsewhere.
There is also a domestic audience. The International Embodied AI Skills Competition drew more than 800 teams to Shanghai this weekend, and Beijing’s new employment guidance explicitly ties job creation to AI development. K3.1 Pro is as much a signal to that home crowd β proof that Chinese labs can build at the frontier β as it is a product for export.
What developers should watch next
The real test starts when independent developers run K3.1 Pro on their own code. Watch for three things: whether the Terminal-Bench score holds up on other coding tests, what the API pricing looks like once the rollout completes, and how the Highspeed edition performs under heavy load. If all three hold, expect Western labs to answer with price cuts or faster models of their own.
For now, Kimi K3.1 Pro is best read as a marker: the frontier is no longer a two-country story told only in California. It runs through Beijing and Shanghai too, and the next model that changes your workflow might come from a lab you have never used.
Sources and Resources
Reporting for this article is based on the daily briefing: AI, Economy, and Global Conflict Daily Briefing, Sunday, October 25, 2026, which covers Moonshot AI’s Kimi K3.1 Pro announcement.
When was Kimi K3.1 Pro unveiled?
Moonshot AI unveiled it in the early morning of October 24, with a phased public rollout running through November 4.
How big is the model?
It uses a 2.8-trillion-parameter architecture, with 104 billion parameters activated per inference pass, and a 1-million-token context window.
Did it really beat GPT-6?
On Terminal-Bench 4.0, Moonshot reports 65.7 points against 58.2 for GPT-6 Astra. Those are company-reported figures, not independent tests.
Where can you use Kimi K3.1 Pro?
At launch it is available in the Kimi app and through Moonshot’s API, including a faster Highspeed edition.
Why does the 1 million token window matter?
It lets the model hold a whole codebase or document set in one prompt, which helps agentic coding across many files.
Conclusion
Kimi K3.1 Pro is a statement of intent from Moonshot AI: a giant, fast, coding-first model aimed squarely at the frontier. The benchmark claims are bold and company-reported, so independent testing will be the judge. But the direction is clear β the AI race now has more than one center, and developers have one more serious option to evaluate.
See also: World Reach Daily