Introduction
I have learned to be suspicious of the phrase "open weights" in a launch announcement. Too many times this year it has meant "open weights, coming soon," or — the one I saw just yesterday with Tencent's Hy4 — "open weights, in the coming days," while the download link quietly goes nowhere. So when Moonshot AI said on October 23 that the full weights of Kimi K3.1 Pro would be released by November 4, I filed it under "believe it when I see the repository."
On November 4, the repository appeared. The weights are real, they are downloadable today, and they are the largest open-weight model anyone has ever published. That is the headline, and it is true. But the size of the download, the license attached to it, and the gap between what the model can do and what most of us can actually run are all part of the story too — and if you build websites or software for a living, those details decide whether this matters to you this week or sometime next year.
What Actually Landed on November 4
Moonshot posted the weights on Hugging Face under the repository name
moonshotai/Kimi-K3.1-Pro — no access gate, no form, no waitlist. The download is 96 safetensors shards totalling about 1.56 terabytes, and Hugging Face's records list the repository as created on November 2, going public during November 4. That timing is the opposite of what the rest of the industry did this week: Tencent launched its flagship Hy4 on the same day and its weights are still a promise. Moonshot promised a date and hit it early. In a year when "open" has become a marketing word, shipping the actual files on the actual date is the whole credibility test, and Moonshot passed it.Kimi K3.1 Pro itself is not brand new. Moonshot announced it on October 23, 2026, as the first model in its new K3.1 series, post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of tasks. Since launch it has been live in Kimi.ai, Kimi Work, Kimi Code, and the Kimi API. The weights release is the second half of that launch: the same model, now downloadable — if your hardware is up to it.
The Numbers: 2.8 Trillion Parameters, 104 Billion at a Time
The headline figure is 2.8 trillion total parameters, which makes K3.1 Pro the largest publicly downloadable open-weight model in existence, ahead of everything else on Hugging Face. But the headline figure is also the one most likely to mislead you, so let me explain what it actually means.
K3.1 Pro is a mixture-of-experts model. It has 896 experts, and only 16 of them activate for any given token — about 104 billion active parameters per token. That is how a 2.8-trillion-parameter model stays servable at all: the total is capacity, not per-token compute. When you compare it with a dense model, the honest comparison is the 104 billion active, not the 2.8 trillion total. It is still an enormous model by any standard. It is just not quite as absurd as the headline suggests.
The context window is 1 million tokens — 1,048,576, to be exact — the same as Kimi K3. That is the number that matters most to me as a working developer. A million tokens means you can hand the model a whole codebase, a long legal document, or a year's worth of analytics and ask it to work across all of it at once. Most models I use daily still cap out well below that, and the difference shows up the moment your task stops fitting in a single chat window.
There is a genuine architectural change under the hood, too. In 24 of the model's 93 layers, the usual full attention mechanism is replaced with a sparse variant called Gated DSA — DeepSeek Sparse Attention — which is what makes those million-token contexts affordable to process. The weights are stored in MXFP4 format with MXFP8 activations, trained with quantization awareness, which is why the whole model fits in 1.56 terabytes instead of something closer to six. And it is natively multimodal: text and image input, through a 401-million-parameter vision encoder called MoonViT-V2.
The License: Open for Most of Us, With a Line Drawn at the Top
Here is where I need you to read carefully, because "open weights" and "open source" are not the same thing, and K3.1 Pro is the first and not the second.
The weights ship under Moonshot's own Kimi K3.1-Pro License, not a standard open-source license like Apache 2.0. It is MIT-style in the basics: you can use, copy, modify, and sell the software. But it has conditions that bite at the top end of the scale. If you or your affiliates run a model-as-a-service business and your aggregate revenue exceeds 20 million US dollars over any consecutive twelve months, you must sign a separate agreement with Moonshot before using the software for any commercial purpose. Very large products — those with more than 100 million monthly active users or more than 20 million dollars in monthly revenue — face additional display requirements.
For almost everyone reading this blog, none of that matters. If you are an independent developer, a small agency, or a WordPress freelancer running a model on your own server or through a small host, the license lets you do what you want. The conditions exist to stop the biggest cloud providers from turning Moonshot's model into a competing hosted service without a deal. That is a reasonable commercial boundary, and it is also a reminder that "open" in 2026 usually means "open, with a moat around the biggest money."
The Benchmarks: Strong, With the Usual Caveats
Moonshot's own numbers put K3.1 Pro near the top of the field. On Terminal-Bench 4.0, a coding-agents benchmark, it scores 65.7 percent — ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent. It reports 94.2 on GPQA-Diamond, 92.1 on BrowseComp, and 79.6 on OSWorld V2.1, the computer-use benchmark. The independent Artificial Analysis Intelligence Index puts it fourth among every model it tracks, at 55.1 — an 11-point jump over Kimi K3.
I will say the same thing about these numbers that I say about every vendor benchmark table: they are Moonshot's own runs, and until independent researchers reproduce them on the public weights, they are encouraging rather than settled. The Artificial Analysis score is the most useful of the bunch precisely because it comes from outside Moonshot, and fourth overall for an open-weight model is a serious result.
The one benchmark result that matters most to me is Terminal-Bench, because it measures what an agent actually does in a terminal — running commands, editing files, recovering from errors — rather than what it writes in a single answer. A model that scores 65.7 there is a model I would trust to do real work on a codebase, with supervision. That is the practical test, and it is the one I will run myself now that I can download the weights and test them against my own tasks.
What It Takes to Run: This Is Not a Laptop Model
Let me be blunt about the practical side, because the download button is the easy part. The weights are 1.56 terabytes, and the simplest vLLM setup needs eight data-center GPUs. That is not a home server and it is not a gaming rig. For almost everyone, the honest answer is: you will not run this model yourself. You will use it hosted — through the Kimi API, Kimi.ai, or a cloud host that started serving it on day zero, including Ollama Cloud, Nebius, and Verda. And that is fine. The value of open weights is not that everyone runs the model in their garage. It is that anyone can inspect it, fine-tune it, and benchmark it honestly without asking Moonshot's permission.
The hosted price is genuinely competitive: 2 dollars per million input tokens and 10 dollars per million output tokens — down from 3 and 15 dollars for Kimi K3, a 33 percent cut on a substantially stronger model. At those prices, a million-token context stops being a luxury and starts being something you can actually afford to use on a real job.
What This Means for WordPress and Elementor Builders
You might be wondering what a 2.8-trillion-parameter Chinese model has to do with your WordPress work. More than you would think, and the connection runs through coding agents.
The single most useful thing about a model like K3.1 Pro for a WordPress developer is the combination of a million-token context and strong agentic coding. Give it your whole theme, your plugin code, and your error logs at once, and it can reason across all of them — find the function that is actually breaking, not just the one the error message points at. That is the kind of task where a long context window turns a clever chatbot into a useful colleague.
The open weights matter for a second reason that is specific to our corner of the web. WordPress sites hold client data, and many agencies cannot send that data to a third-party API for contractual or privacy reasons. A downloadable model — even one you run through a private host rather than your own hardware — gives you an option that a closed API never will: the model runs where you say it runs, on data that never leaves your infrastructure. For client work under strict data rules, that is not a nice-to-have. It is the difference between using AI on a project and not using it at all.
What I Would Do This Week
If you want to try K3.1 Pro sensibly, here is the short list I would actually follow.
First, do not download 1.56 terabytes on a whim. Try it hosted first — on Kimi.ai, through the API, or through a cloud host. Run it on your own messy work: a real codebase, a real debugging session, a document you actually need summarised. Vendor benchmarks are a starting point; your own tasks are the test that counts.
Second, if you run a WordPress agency, look at the long-context coding angle before anything else. Feed it a staging site's theme and plugin code and ask it to audit the lot. Compare its answers against a model you already trust, on the same prompts, and keep score yourself.
Third, read the license before you build a product on it. For personal use, small teams, and most agencies, you are fine. If you are building a hosted service that might cross the 20-million-dollar revenue line, talk to a lawyer first — or just use the API and let Moonshot handle the licensing.
Conclusion
Kimi K3.1 Pro's weights release is the rare AI story this week where the headline and the reality line up. Moonshot promised a date, hit it early, and published the largest open-weight model ever released — 2.8 trillion parameters, a million-token context, strong coding benchmarks, and a price cut on the hosted API. The license has a commercial boundary at the top end, the hardware requirements put self-hosting out of reach for almost everyone, and the benchmarks are still vendor-reported. None of that takes away from what actually happened: the files are on Hugging Face, ungated, and anyone can download them today.
FAQ
Are Kimi K3.1 Pro's weights really public?
Yes. Moonshot published them on Hugging Face as
moonshotai/Kimi-K3.1-Pro on November 4, 2026. The repository is ungated — no form, no waitlist — and holds 96 weight files totalling about 1.56 TB.How big is Kimi K3.1 Pro, really?
It has 2.8 trillion total parameters, with 104 billion active per token. It is a mixture-of-experts model with 896 experts, 16 of which activate for each token. The honest comparison with dense models is the 104 billion active figure.
Is it really open source?
Not in the strict sense. The weights ship under Moonshot's own Kimi K3.1-Pro License, not a standard open-source license. You can use, copy, modify, and sell it, but model-as-a-service businesses earning over $20 million a year need a separate agreement with Moonshot, and very large products face additional display requirements.
How does it compare with Tencent's Hy4, launched the same week?
Hy4 is a 770-billion-parameter model with vision and computer-use features, but its weights have not been released yet — Tencent says they are coming in the following days. Kimi K3.1 Pro is larger, and its weights are downloadable today. That is the practical difference.
Should I use it for my WordPress work?
Try it hosted first, on real tasks: a long-context code audit, a debugging session across a whole theme, or a document review. Its million-token context and strong Terminal-Bench score make it a serious coding agent. Keep client data rules in mind, and read the license before building a product on it.
