There’s no recall button for a model file. Moonshot AI posted the Kimi K3.1 Pro open weights to Hugging Face around Nov 4. From then on, anyone with enough hardware can keep a permanent copy of a 2.8-trillion-parameter model.
Moonshot says it made that copy safer before letting go. It pulled cyber-related data out of training, and it says the model still trails the closed frontier on hacking skills. Nobody outside the company has checked those claims yet.
That gap, between a lab’s promise and anyone else’s proof, is the real story. The specs matter, and they’re impressive. But this release forces an older, harder question: once a frontier-class model leaves the building, what does “safe” actually mean?
Kimi K3.1 Pro open weights: what Moonshot actually released
The files live in the repo moonshotai/Kimi-K3.1-Pro, and the official model card lays out the basics.
Kimi K3.1 Pro uses a mixture-of-experts (MoE) design. The network holds 896 specialist sub-networks, called experts, but wakes only a handful for each token. Per the card, it picks 16 experts plus 2 shared ones every time. So the model has 2.8 trillion parameters in total, yet only about 104 billion do any work on a given token.
The rest of the spec sheet, from the card:
- 93 layers and a context window of 1,048,576 tokens, roughly a million.
- Text and image input, through a 401-million-parameter vision encoder called MoonViT-V2.
- MXFP4 weights and MXFP8 activations. Moonshot trained the model with these low-precision number formats in mind (quantization-aware training), which keeps the files smaller.
- Reasoning effort set to “max” by default. Moonshot lists vLLM, SGLang and TokenSpeed for serving, and the API name is kimi-k3.1-pro.
K3.1 Pro is a post-trained version of Kimi K3, which Moonshot opened up in July. TechNode reports that the update adds expanded reinforcement learning for coding, 3D creation, financial research and long-running agent tasks.
Moonshot’s own framing was upbeat. Its official X post, as RuntimeWire quoted it, announced it was “Releasing the model weights of Kimi K3.1 Pro” and promised to “keep advancing open intelligence.”
What’s still unconfirmed
Several details don’t line up across sources. Treat these as open:
- The exact date. RuntimeWire puts it at Nov 4, citing Moonshot’s X post. TechNode reported it on Nov 5. Hugging Face metadata shows the repo created on Nov 2, but that’s a creation stamp, not a launch date.
- “3T-class” vs 2.8T. The card calls this “our most capable 3T-class model to date.” The actual count is 2.8 trillion. Read “3T-class” as a size bracket, not a figure.
- A “Highspeed” variant. Moonshot’s launch post reportedly mentions one. The model card doesn’t, and nobody has confirmed weights for it.
- Benchmarks and hardware. Every score comes from Moonshot. No third party has reproduced them yet, and the card gives no minimum hardware.
Open weights vs open source: what the Kimi license really allows
Get this straight first. Kimi K3.1 Pro’s open weights don’t make it open source in the strict sense. The model ships under a custom “Kimi K3.1 Pro License,” not an OSI-approved open-source license. RuntimeWire noted that the license limits how open the release really is.
Open weights means you can download the trained model and run it. It doesn’t automatically hand you the training data. And it doesn’t mean the terms come without strings.
The Hugging Face page doesn’t summarize those terms. The clearest reading so far comes from the Korean outlet braindetox, which went through the LICENSE file clause by clause:
- Clause 2: companies selling the model as a service need a separate agreement with Moonshot once their combined revenue tops US$20 million over any 12 consecutive months.
- Clause 3: commercial products with more than 100 million monthly active users, or more than US$20 million in monthly revenue, must prominently display “Kimi K3.1-Pro.”
- Clause 4: internal use, and use through Moonshot’s official products or certified partners, falls outside those conditions.
Our read? For a university lab or a small startup, this behaves a lot like open. The conditions only bite at serious scale. They look aimed at large cloud resellers who’d otherwise profit from Moonshot’s work without paying or crediting it.
That’s a commercial reading of a secondary source, though, not legal advice. If you plan to build on the model, read the LICENSE yourself.
Why a 2.8-trillion-parameter model is open to few people in practice
“Open” assumes you can actually open the thing. Downloading the Kimi K3.1 Pro open weights is the easy part. The repo holds 96 safetensors shards that add up to about 1.56 TB, or 1,561,293,160,568 bytes by braindetox’s count.
Moonshot doesn’t publish a hardware requirement, so here’s our own back-of-envelope math. Divide 1.56 TB by the memory on an 80 GB data-center GPU and you get about 20 cards. That’s just to hold the weights. It leaves nothing for a million-token context.
So who is this open to? Cloud providers, well-funded labs, governments and large companies. Hobbyists can download the files. Very few can run them.
For everyone else, the practical route is still an API. That shifts the meaning of the release. It’s less a gift to individuals and more a move in an ecosystem game.
TechNode reports that Huawei announced inference support through vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 hardware. A Chinese frontier-class model, ready to run on Chinese chips on day one, says something about supply chains as much as software. It also complicates the tidy scoreboard in debates like the one we covered in US China AI Gap: Is It Really Down to Just 3 Percent?
Stripping cyber data: does it hold once the weights are public?
This is the part of the release we find most interesting. It’s also the thinnest on sourcing. According to Moonshot’s own statements, as relayed by braindetox, the company:
- removed cyber-related data from both training stages while keeping general coding ability;
- tested the effect on xdev-bench-hard, an internal benchmark, where a checkpoint from before the removal solved 90.5% of tasks (pass@3) and the final model solved 81.0%;
- says the model’s cyber capability stays behind proprietary frontier models that have been on the market for at least three months;
- says it aims not to release open weights that extend the absolute cyber frontier.
A caveat before going further. In the sources we checked, only braindetox carries this policy text. TechNode and RuntimeWire don’t mention it. The Hugging Face card makes no explicit safety statement; it just lists scores. And Moonshot assessed every number above itself.
What the numbers do and don’t show
Start with the headline drop. Falling from 90.5% to 81.0% is real, but modest. On Moonshot’s own hardest internal cyber test, the shipped model still solves about four in five tasks.
The public card scores tell a similar story. K3.1 Pro scores 74.7 on ExploitBench, against 83.0 for Claude Fable 5.1 in the card’s own table. braindetox’s table lists Claude Opus 5.5 at 91.0 and GPT-6 Astra at 100 on the same test. On ExploitGym, the card shows 20.7. So yes, it trails the closed leaders. It doesn’t trail by miles.
Then there’s the deeper problem. Removing training data is a pre-release safeguard. Once the Kimi K3.1 Pro open weights sit on thousands of drives and anyone can fine-tune them, Moonshot has no post-release control at all. Someone with enough compute can try to train the missing skills back in. Whether anyone succeeds is unknown. The point is that nobody can stop them from trying.
Who sets the bar, and who checks it?
The “behind the frontier” standard has a quirk, too. It defines safety relative to closed models. As closed labs get more capable, the bar Moonshot has to clear rises with them. That’s a sensible policy for a lab racing its peers. It’s a strange one if you care about absolute risk.
There is one outside data point, for the predecessor. When Moonshot released base K3 in July, The Decoder reported that an independent test by the UK’s Cyber Institute found K3’s cyber skills trailed frontier models by a wide margin. No comparable outside test of K3.1 Pro has surfaced yet. That’s exactly the hole independent red-teaming exists to fill, and no lab can fill it by grading itself.
What Kimi K3.1 Pro means for the AI race
The benchmark picture is mixed, and every number comes from Moonshot, run at max effort. Here’s how K3.1 Pro stacks up against Claude Opus 5.5 and GPT-6 Astra in braindetox’s table:
- Terminal-Bench 4.0: 65.7, vs 66.4 (Opus 5.5) and 58.2 (GPT-6 Astra)
- SWE Marathon v1.1: 51.7, vs 46.3 and 42.5
- HLE-Full with tools: 70.0, vs 67.7 and 57.2
- Program Bench: 31.5, vs 65.0 and 50.0
- HLE-Diamond: 39.4, vs 60.9 and 65.8
Read that slowly. K3.1 Pro runs level or ahead on long agentic coding and tool-assisted reasoning. It falls far behind on Program Bench and HLE-Diamond. Test conditions differ, some sets are internal, and nobody has reproduced any of it.
The card adds strong numbers elsewhere: 92.1 on BrowseComp, 94.2 on GPQA-Diamond, 79.6 on OSWorld V2.1 and 84.1 on MMMU-Pro.
The DeepSeek precedent
We’ve seen this pattern before. A Chinese open model crowding the closed frontier was the whole story when DeepSeek V4 arrived. With base K3, Moonshot claimed “2.5 times more intelligence per unit of compute.” It also open-sourced attention kernels, an MoE communications library and agent tooling, The Decoder reported. The same article floated, as speculation, that Moonshot might lean on distillation. That stays an open question.
Our take: the race story isn’t simply “China caught up.” The Kimi K3.1 Pro open weights put near-frontier agentic coding in a downloadable file, with domestic hardware ready to serve it. That squeezes closed labs. It squeezes safety norms harder.
If every open release promises only to trail the frontier by a few months, “a few months behind” quietly becomes the industry’s safety policy. Nobody voted for that. For where this leaves the biggest players, see our breakdown of the AGI development race in 2026.
What to watch next
- An independent cyber evaluation. The UK test of base K3 set the precedent. A similar test of K3.1 Pro would turn Moonshot’s claims into evidence, or not.
- Third-party benchmark runs, especially on Program Bench and HLE-Diamond, where the gap looks widest.
- Fine-tuned derivatives. Watch what people build on the weights, and whether anyone tries to restore what Moonshot removed.
- The Highspeed variant. If its weights appear, check whether the same safety claims cover it.
- A legal reading of the license from someone qualified to give one.
Open, capable, and still unverified
We think Moonshot deserves some credit here. It published a before-and-after number on its own safety work, and that number doesn’t flatter it much.
But a promise graded by the company that made it is still a promise. Until outside testers get their hands on this model, the most honest label for the Kimi K3.1 Pro open weights is the one in this heading.
If you’re thinking of building on it, start with the LICENSE file and the model card, not the launch post. And ask yourself the question Moonshot can’t answer alone: who checks the lab once the files are already out?