> First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models. Second, we intend to progressively raise the cyber capability available in open weights.
So they admit that they could do more, but they won't?
> So they admit that they could do more, but they won't?
Yeah, there's another paragraph:
> Identifying and removing cyber-related data. During development, we observed vulnerability research performance improving steadily as post-training scaled. Tracing the source of this trend, we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix — a natural byproduct of training on broad data. In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness. We will continue to apply appropriate controls at every training stage, of which cyber-related data identification and restriction is one example.
(emphasis mine)
It's clear that cyber is a touchy subject right now. But it's also clear that RL works on it, because it's highly verifiable. So smaller models are going to become very interesting, because they can be fine-tuned later to cover this field.
I think this is misguided. Considering how much code is AI written these days do we really want LLMs writing insecure code because we lobotomised them?
It’s not much of a deterrent to a motivated organization.
They are basically saying that models themselves are already capable. If you fine tune them with some cyber security tasks, that should make them potent. :-/
The "rule" here is, if you read the disclosure carefully, that they have cyber trained models in house. But you do not get them, you only get the non-cyber models.
With frontier competitors and the U.S. govt actively pushing "concerns" about Chinese models, Kimi including this statement on "offensive cyber capability" seems like smart defense. Models are opaque and training details are trade secrets, so whatever they did might be negligible or significant and no one can tell. This bit is also artfully done:
> "below what our pipeline can produce, and remains behind that of frontier proprietary models."
Like Ant and OAI, they've realized vaguely claiming to have nerfed their latest model to an unspecified degree not only looks 'good guy' prudent, it provides a built-in excuse when a model fails.
Or it's a cop out. I could totally outrun Usain Bolt, but by design I don't push to set new World Records. I will increasingly show faster times though.
> Or it's a cop out. I could totally outrun Usain Bolt, but by design I don't push to set new World Records.
The benchmarks show Kimi K3.1 Pro outrunning practically all models. It's weird to hear someone like you accusing them of making up excuses for not lagging behind in discussions on how they are ahead of the pack.
Well, in the publication they claim to have identified specific cybersecurity post training tasks that nerfed capabilities when they removed them, which seems like a reasonable methodology.
Seeing as no one besides blackhats want frontier cybercrime capability to be open source, this seems like a very sensible thing to do.
So they have published a reasonable means, and they have a motive. I'd put my money on "probably true".
> Seeing as no one besides blackhats want frontier cybercrime capability to be open source
Sure. The problem is that there is no practical difference between "cybercrime capability" and "cybercrime defense capability."
All of my PRs go through an adversarial security review by GLM-5.3, because it's the only smart-enough model with a good understanding of cybersecurity issues that won't regularly refuse the job. Removing that ability from frontier models helps nobody except cyber criminals.
I am a defender, and I really really want to have access to high quality models that are good at finding vulns and hacking. Leaving them in the hands only of government agencies and orgs with enough resources to do extensive training / fine tuning seems like the absolute worst possible case for the future security of global software systems.
The problem is, as anyone trying to address security concerns in claude can attest to, the guardrails trigger even when trying to address real bugs. Good cyber defense requires an understanding of the offense.
How you avoid the "fix my code" attack, the code being someone else's code, I do not know. But nerfing the model doesn't seem like the best path forward either.
This is not about guardrails, it's just not post-training the model on tasks that specifically relate to cyber-offense. It's a very lightweight intervention, and it makes the model less of a target for bad actors.
> Seeing as no one besides blackhats want frontier cybercrime capability to be open source
I'm not interested in black hat anything, but I am very invested in making sure the products I work on are not compromised. Given that state agencies have "frontier cybercrime capability", I would like to have the same.
Please don't change "cybersecurity" to "cybercrime" from one sentence to the next as a rhetorical flourish. It comes off as disingenuous and makes it harder to get your real point across.
Yeah that's a huge downside. That the model does the task without tripping over its own feet, like discovering a vulnerability the model itself accidentally built 15 minutes ago and triggering its own guardrails, made the Chinese models attractive to use for security-adjacent things. Also useful to red-team your own stuff. While I can understand the reasoning behind it (and I guess having a model that's really good at both offense and defense released into the ether could lead to CCP asking some very pointed questions if Moonshot managed to get blamed for some incident stemming from their model's misuse) it's still a huge disappointment.
I don't want to give US labs the results of my DNA tests and my family tree going back to the 16th century just to secure my shitty app, but if the Chinese are bowing out as well it seems the only alternative is to admit that some script kiddie with a bunch of tokens will just pwn you now and then.
If the Chinese want it, they have it. BGI does an enormous amount of sequencing. My own genome is entirely public (or was, and will be again when/if Personal Genome Project comes back) so you can watch and see if it harms me.
> I don't want to give US labs the results of my DNA tests
My friend, you realize that if you have had blood drawn for a genetic test, there is a very decent probability that the sequencing was done by a Chinese lab operating in the US, right? The Chinese probably have more US genomes that the US does.
Also, illumina used to ship products with hardcoded S3 access, so purchasers of used equipment could gain access to sequencing information generated by the original owners. Too many links to cover all of it, here's the Gemini summary:
Moonshot (Kimi) seems to be unashamedly into American culture[0], I wonder if they're paying closer attention to and giving more credence to some cultural aspects of American AI discourse including aspects of AI caution/panic.
All postwar english language culture that isn't much more explicitly non-American than that is American. All good rock music is American by default, unless it is near-constantly screaming otherwise.
Regardless, expand both sides of the point from "American" to whatever you need to, the idea is the same either way.
That just conveys an idea of being extremely desperate for being the leaders when you're clearly not it, and that despair is so great that you don't mind switching teams midway through just to be able to claim you're winning.
If we had a technique that would defeat almost every lock in the world, we might be prudent to delay releasing that until people have had time to change the locks.
Same general benefit as coordinated disclosure, except it's a big game-theory problem where we're likely to eventually have sufficiently capable defectors.
I mean they say they will release the weights by November 4, so presumably after that it won’t be hard for people to fine-tune it to be better at cyber right?
Presumably if it knows how to find and fix bugs it can be told how to exploit them. But who knows what it will actually be capable of doing in the real world as opposed to Kiki’s benchmarks
It's not actively nerfed like the Western frontier models, it's just downweighting offensive "cyber" as a domain for post-training/fine tuning. Which, if anything, ought to improve model capabilities wrt. other domains.
> When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M.
This is 100M fewer than Opus 5.5, 40M fewer than Fable 5.1, and about a third of Sonnet 5.5. In the top-6, it's 3rd, beaten by Muse Wonder at 140M and Astra at, impressively, just 60M.
That's versus Opus 5.5 max effort, no? Opus high has the closest cost and intelligence characteristics and generates far fewer output tokens per task than the one Kimi 3.1 Pro effort level we have data for.
Yes, max effort is what the displayed-by-default score on the index is measured by. At xhigh, Opus 5.5 outputs 100M tokens, but remains only 1 point ahead of K3.1 rather than 5. And Sonnet at xhigh outputs 110M tokens and trails K3.1 by 3 points rather than leading by 3, at the same price point per mtok.
Opus 5.5 high trails K3.1 by 1 point, but outputs only 53M tokens, so it indeed comes out cheaper per-task. I think it's still remarkable that this is effectively an open model that matches the frontier for the first time in a long time, pretty much since DeepSeek R1.
> When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M.
So their base price is 1/5th of Opus 5.5 per token, but if you want "highspeed" it's 2x that and then I guess it generates 2x the number of tokens as well?
That might suggest that Opus is competitively priced after all.
In an agentic pipeline, the output text is 2-5% of the total tokens used, so verbosity is a nonissue costwise. Although, higher verbosity could cause input to balloon faster and potentially cost more, but that would be captured by the cost-per-task metric.
But as noted by other replies, it's still lower than Opus/Sonnet.
I don't trust these Artificial Analysis indexes and I'm not sure why anyone considers them authoritative, the site just sprung up a few months ago and now everyone quotes it?
From my own use of the models I find their ranking to be totally inaccurate.
Do you have an alternative? Its a benchmark collection. When a benchmark gets saturated, they remove it from the index and add new ones.
I think lm-arena is a better concept that reflects the actual target users, but it only captures chat, and relies on users spending their time improving it.
https://epoch.ai/eci?view=graph&tab=release-date is at least an alternative. They are slower to update, but entirely anecdotally I find the numbers are aligned with my personal observations a bit better than AA. But I am also highly aware this is some pretty confirmation-bias heavy ares of meta-benchmark picking...
They are the only site that really combines all benchmarks into a single place. And their intelligence index is pretty broad, including knowledge work and finance tasks as well as coding, so if your use case is narrow it may not match the general benchmarks
It's appealing to me because it's not just one benchmark but a summary of a collection of benchmarks. Also the current rankings have Opus on top by far and everyone's experience is that Opus is best by far so it's not massively inaccurate
Moonshot has been frontier since this summer. At this point whenever they release a model it's on par with whatever the latest OpenAI/Anthropic model release was.
Not sure I'd put Google there in practice. And while Muse is an impressive jump from behind it's still not quite at the frontier either IMO.
You either include all labs which can reasonably produce models which are capable enough to be considered generally on-par with the best available models, in which case there has been way more than 2 frontier labs for a while now, or you take a more literal definition and only include labs which are capable of producing models which are considered "the best" out there, in which case there are basically just 2.
I think the former definition is a bit more meaningful since it avoids issues with determining what actually constitutes "the best" model out there. In that case, Google has been a frontier lab for a long time, Meta arguably as well, and you're also ignoring xAI which has been competitive for a while now. And, obviously, there's a bunch of Chinese labs which would deserve the title as well.
But if one is being strict (and willing to accept generally accepted definitions/measures/etc.), then there has only ever been two frontier labs: OpenAI and Anthropic. They're the only labs to have had "the best" models at any point in time since LLMs became a thing.
right now it looks like only anthropic since openai made the staggering business decision to release their midrange model (terra) rebranded as sol 6.1 and quantized astra into the ground.
everyone else mostly seems to have just decided to cheat on benchmarks to make their models look better than they actually are, and none seem to want to actually try to make better models.
I thought it was 6 Sol which was a rebadged Terra? You think 6.1 Sol is too? It does seem to benchmark better than 5.6 Terra? Regardless agree both models have been very bad for openAI's Sol brand. To be fair it's at least cheap compared to other OpenAI models.
It's still just 2. Google hasn't been at the top in a long while, and even when it was, it was just a tiny improvement over what other labs had previously released. Meta has never been at the top. Only two labs release LLMs that really push the limits out.
I'm not sure what counts anymore. You've got OpenAI and Anthropic; then Meta, Google, and X; then over to Alibaba/Qwen, Z/GLM, Moonshot/Kimi, Deepseek; and Mistral in Europe. How many people out there in the real world can actually come up with tasks that differentiate those 10?
Can someone please tell me what benchmarks to look for that measure world knowledge? For sspitballing, critical analysis, business projects, etc. Is the model size the only way to measure it?
I've not tried it out, but in the past I've been pretty disappointed in the performance of Kimi models. Deepseek, Qwen, and GLM all appear to perform much better even though Kimi often has higher benchmark scores. Ditto for Mimo.
The older Kimi K models (2.5 and 2.6) weren't bad for the time, but all of their models have had problems with overthinking. This also makes them significantly more expensive in practice than their pricing suggests, since they produce way more output (reasoning) tokens than most other models. Anecdotally, I've also had problems with all the pre-K3 models thinking themselves in circles trying to solve very simple problems that other models didn't struggle with. (I haven't played with K3 much because it's just too expensive.)
I have the opposite impression. I find K3 to be extremely capable at full-stack work; its UI design skills are great. I'd love to use it more, but the quota on Kimi's plans is abysmal. I wonder if these new models improve that.
If you are expecting any sort of usable limits on their subscription plan I suggest adjusting your expectations. I subscribed to $99 a month plan and within 47 minutes with one session (not asking for swarms or anything) was limited by 5-hour quota.
Same experience. I had the $19 monthly plan for experimenting and running benchmarks, but it was a uselessly small amount of usage, a single task would blast through the whole five hour limit and a big chunk of the weekly limit. All of the US provider plans are remarkably more generous (though OpenAI plans have gotten a lot stingier lately, where they used to be the most generous).
Had a similar experience here. Their limits are very, very low, and ultimately make it not worth subscribing. Can't leverage a good modal if it exhausts its 5-hour budget in 30 minutes.
Paying API prices for Deepseek 4.1 Flash feels like a much better deal at the moment.
Welp, that ruins this news for me. I'm glad open models are catching up but if I can't practically run them locally _or_ through a flat fee then it doesn't really help me. I fear API pricing.
This'll end up on opencode go soon enough. $40/month on there gives you a bit more than just "light use"
Don't use frontier models like this for problems that can be solved with flash models, and you'll see API pricing isn't that terrible. Still cheaper than anything with remotely this level of capability 2.5 months ago.
Yeah, the plan is far too expensive for what it provides. I wonder if the quota with these new models will be better, but I'll wait for others to try this out before wasting another hundred bucks.
Yes, if you just look at API pricing, the cost is about on par with its capabilities. The problem is that we all use subscriptions, and when you compare subscription pricing, Kimi is much pricier than pretty much all of their competitors, including Chinese ones.
Fascinating that we’re converging on AI safety and alignment in time across jurisdictions. The cybersecurity approach taken here is interesting because should the model be opened it’s still not dealignable because it just doesn’t know enough.
The slow warming of the water is a good approach. I’ve been using abliteration.ai’s cybersecurity models to pentest my own software and it produces much more usable defensive results. Opus 5.5 can fix everything but the guardrails trigger too easily.
The Fable ban affair was a nightmare for defensive cybersecurity. As far as I’m concerned we’re already in the world of Void Star and Cloudbreaker is here.
> should the model be opened it’s still not dealignable because it just doesn’t know enough
This is the only AI-safe approach - see the OpenAI agent swarm incident where the swarm engaged in offensive cyber because guardrails were not active for their task and the model had clearly been post-trained on that narrow domain.
Does this make sense? It is the same as when Stable Diffusion tried to not have "people laying down" in their training data in order to block the generation of sex pictures.
Deliberately keeping all models ignorant doesn't make sense. It won't work (whoever breaks the pact has a massive advantage); it won't be enforceable; it will have downstream impacts on things you don't want it to be ignorant about; this kind of regulatory regime invites adding more and more to the "no-no list".
Is it not possible to focus on hardening? And on new legal liabilities for operators who do security attacks?
I bought a yearly subscription to kimi when it was the hype and never use it. The 5h limit makes it completely unusable. You barely start doing anything and you hit the limit immediately. I advise everyone against using kimi.
Ran it on my adversarial programming language (KillSwitch) benchmark[1]. Very significant improvement over K3. On par with Gemini 4 Argon and Fable 5.1. Much slower than both though.
But this is now the highest scoring open-weight model. Amazing stuff.
There is a ton of questions to ask regarding their removal of cybersecurity data.
First, how much “cybersecurity data” is just standard software knowledge that one can derive cyber security exploits from?
Did they also get rid of all references to hacking in general?
Are they implying that all the recent exploits are just things that were already in the data?
For a model it doesn’t seem that far fetched to make the leap from “I have the knowledge of what hacking is” and “I know how computer systems work” to just being able to exploit stuff anyways.
If you're an amazing, top 1% software engineer, and you switch jobs to become a penetration tester, you're not going to be top 1% day one. There's knowledge specific to penetration testing you'll need to research and absorb and practice before you're good.
For the models, it's always day one. They don't learn from experience. Whatever they know after training is all they'll ever know. So if you don't teach them penetration testing in training, they won't know it. Maybe they could theoretically derive everything from first principles, and so could a top 1% software engineer, but that's a very slow and inefficient way to do things.
> For a model it doesn’t seem that far fetched to make the leap from “I have the knowledge of what hacking is” and “I know how computer systems work” to just being able to exploit stuff anyways.
You'd be surprised. Cyber offense is a very narrow domain, especially given the extensive mitigations that are being used in modern systems. Leaving it out of post-training tasks unless you're specifically interested in it is the no-brainer approach.
Will there be a book of secret knowledge that only the initiated will be allowed to see? And they will all swear an oath to never input it into an AI system?
And then we'd also do that for all kinds of potential terrorism, sex-crimes, political subversion, and evil advocacy?
It would be the "restricted section" like in the Hogwarts library?
Post-training is not about "secret knowledge": it's literally about training the model in a RL environment on very specific tasks (including, potentially, cyber exploitation). It's just as compute-intensive as pre-training these days, if not more (given that several checkpoints of the same pre-trained model are being released, showing sizeable gains in real-world capabilities for non-trivial productivity tasks).
Perhaps the gap is just too wide, but to use a more simple analogy. I see this like someone knowing everything about biology, simple mechanics, metallurgy, and human history but never hearing about an axe then assuming that person would never come up with the concept, use case of, or means to create an axe.
In looking into what Anthropic has accused the Chinese AI companies of, "distillation" is not the crux of the matter: they accuse them of the use of massive numbers of breached computers and credit cards. Supposedly (though not explicitly demonstrated) distillation campaigns used "dark web" resources (like stolen and faked credit cards and p0wned IOT devices, routers, etc) that were used to create many many accounts for distillation. (Note: I am generally in favor of distillation - it is "normal" use as far as I can see.)
So there are these "resources" out there: massive botnets, stolen credit cards and credentials, etc. These often seem to be linked to or supported by state actors.
How can we hope that these "standing reserves" aren't used generally by AIs as if they are just another "skill" they know how to leverage?
I imagine that Chinese AI labs would be more than happy to pay for tokens to train on. If Anthropic goes and bans them as quickly as it can identify to them, of course they'll shift to using lots of cheaply obtained sockpuppets to keep getting the same data.
I doubt that an AI could do anything with these that legions of scammers haven't already figured out how to do - after all, it's not like AI set up this distillation setup in the first place :)
Hardening against what? Apologies if I am being dense here.
The way I see it, automated systems should basicially have access to the financial system anyways. This is some of the appeal of cryptocurrency.
Sybil resistance built into the traditional financial system (for example firewalling based on residential IP, requiring banking apps, etc.) is a massive annoyance that can obviously by circumvented at scale.
You think it is a lost cause to get credit cards, for example, less steal-able? Apparently there is a way to create fake cards that work for a few minutes, and stolen cards can be used. A lot of the current security is inelegant and lame. Could data-breech laws not light a fire under some company's asses?
Hardening in this case could mean biometrics, or it could mean rethinking the trust model so that controlling an army of duplicates confers no advantage because you have to be trusted in order to do any damage anyhow. Containing multitudes buys you nothing.
I hope we chose the latter. Biometrics are creepy.
I'm anti-Anthropic and think their accusations are total PR FUD-packing. They don't focus much on the cc fraud because their PR firm landed on pushing the (tech-sounding) word "distillation" to try to scare people in the US and to try to make it a geopolitical issue where trade negotiators can put pressure. The Chinese companies see an advantage in being "open" right now, and it is working. They will not stay open once they see an advantage there. (Contrast with GNU.)
But I've slowly come to realize that AI+"dark resources"(scare quotes)=something bad! And this is imminent (maybe I'm naive and it is already ubiquitous).
We see here, with Kimi, that there's growing traction with the idea of eliminating types of training to avoid evil. I'm skeptical of that.
I really like K3, I still regularly consult it alongside Astra and Opus/Fable.
It is spiky in different places than them and can be surprising in things it gets or catches they don't or don't in the same way. And it has a default personality that I like.
Is highspeed a different model like a flash model or is that just a api/platform feature thats a normal fast mode? I can't tell from the announcement but I might just be missing something obvious?
Interesting that they specifically nerf cyper-capabilities of open weight models:
"In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
It's a bit fascinating to me how all of these labs have rolled their own (it sounds like) completely unique RL environments.
I wonder how much the industry will centralize around a few key patterns, and if we'll start to see some RL framework projects gain traction and be used by the bigger labs
Using it now via API (3.1 Pro on Open Router) in a HITL setup (Cline in VSCode).
Off-hand it's quite good. I've been using K3 for the past few months almost exclusively for coding work and have been stunned by how good it is. Excited to see how far 3.1 Pro can be pushed.
It would be super useful to have more specific analyses of what areas the model is good at. I'd like to know when I should reach for this model instead of Anthropic's or OpenAI's.
For example, Opus 5.5 is better at UI design than Sol. Anyone who has used both models knows that, but it's not something you can look up anywhere.
I'm sure that there's a need for this kind of benchmarking, not just for coding.
One things Kimi is useful for is accessing/querying websites that have paywalls or try to prevent use by AI agents. It is aggressive and persistent for that, though not always successful.
Not sure if this is benchmarkable besides for the own person deciding which one they like more. Models difference is more about taste, preferences and prompt style.
Furthermore, "UI design" is not a narrow enough category, I think. Sure, statistically significant differences in performance between models do manifest when benchmarking against "UI design", but there's still just so much variance within the category. Opus 5.5 can have excellent one-shot results, but totally shit the bed when it's placed in an existing codebase and should uphold the UI/UX conventions, design language, and existing component primitives already established in an app.
But that's just the obvious. There are task by task differences (component by component, page by page, context by context) that the model is worse and better at, even though they are all within the general category of UI design.
I honestly don't have the clarity of mind to explain this coherently, but the crux of my point is that there's a huge variation in it's performance in different contexts, often ones that seem near similar, and in result, even Sol can beat Opus 5.5 in some variations of UI design.
And this is my problem in benchmarking and comparing model performance in general. Presently, we try to create broad categories to benchmark against, create a set of tasks, and then average out the results. That works generally, but it's kind of misleading and it's partly what drives people to feel a disconnect between benchmarks numbers and real world usage on their specific tasks.
It's such a messy field that it's kind of maddening.
I had a friend yesterday tell me he likes the way Claude writes and finds it easier to understand compared to Gemini. But he mostly doesn't use Claude for coding. Whereas I find Claude insufferable in writing often and prefer Gemini for writing. But I don't use Gemini for coding.
Though to be fair Opus 5.5 is better than Opus 5 was in that regard. I still prefer the OpenAI models to either of them when it comes to regular prose, though.
Extrapolating a bit, Moonshot AI provides Kimi K3 at $15 / $3 / $0.3 per million output/input/cached tokens. A 30% reduction would put K3.1 at roughly $12 / $2.4 / $0.24.
So Kimi K3.1 Pro would be generally competitive with the latest Sol and Opus models on API pricing. And that's before factoring in third-party providers, some of which serve Kimi K3 for significantly less than Moonshot. For example, inference.net offers K3 for $10.80 / $0.239 / $0.21.
(One caveat: A Claude or GPT subscription still gives you way more value compared to API pricing.)
Exciting times in any case! We really need more direct competition to OpenAI and Anthropic.
To be fair, some of the benchmarks they highlight might not represent the reality, for example they say:
> On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).
> Our biggest finding is that Terminal Bench 4 is closer to solved than its leaderboard suggests. In 41 of the 65 tasks we reviewed, models could do honest, correct work and still fail, because the grader checked a rule the instructions never clearly stated. We also caught frontier models cheating their way to full marks.
> With the same settings as on Harbor, Claude Sonnet 5.5 went from passing 62.8% of its attempts on these 65 tasks to 89.2%. The best score on Harbor's leaderboard for the same tasks is 65.8%, by Claude Opus 5.5.
To be fair I have a feeling that if you added 'animated' to the SVG request, the output quality from many other models would improve, just because of the extra chain-of-thought work needed to fulfill the more complex request. IMHO the animation requirement should be added to the canonical pelican prompt.
definitely not frontier though, as even on max both legs are in front of the frame; although I am wondering how much getting that right correlates with model intelligence.
The biggest (of many) failing of the pelican test is that it's only one prompt - what matters most these days is how well a model can drive a lengthy agentic tool calling session.
I've consistently heard the opposite from Chinese programmers. Maybe the performance is better in some tasks or maybe the sample size of San Francisco is just skewed.
SWE2 is post trained apparently on Kimi 3. Using it now on Devin, as it's free until the end of the month for a pro account (hey, $20 a month), and it's pretty strong. If I had to pay for tokens I'm not sure how efficient it would be, but definitely Kimi 3 seems a great base. I co-code with ChatGPT 6 and between the three of us it's pretty effective.
So interested to see what the Kimi 3.1 has to offer.
> For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness. We will continue to apply appropriate controls at every training stage, of which cyber-related data identification and restriction is one example.
I like it. If the increase in unrestricted cyber capabilities can be slowed down, please be it. The world needs to get ready, and that takes time. Way more time than even a slow progress would allow.
Unless you want to see more of what happened the past few days.
I think the argument that "cyber capabilities should be democratized, thus keep producing them now, fast, and for everybody" is a very naive one.
Great idea for handling their cyber, by nerfing the model specifically for cyber through removing cyber data from pretraning. This is a way better approach than refusal, which can be trivially abliterated out.
If someone wants to use this model for hacking they will have to train the 3T weights to do so, at that point it's better to use GLM 5.3, so by making it economically unviable to use their models for offensive cyber Kimi has effectively made themselves immune to being targeted in the cyber regulation push efforts.
This is what *actually* pacing the (cyber) frontier looks like.
looks extremely impressive, the full benchmark table isn't opening for some reason though so do we really have agi if small important things like that keep breaking?
first time? Welcome to the irrational SV investor world. Its about hype, perception, politics, perceived future returns, not actual market fit and the proven viability of a technology. Its pets.com, myspace, and NVIDIA pumps because people want to 10x the money now.
Everyone and their dog has been saying its a bubble, there is no moat, this is all known techniques being scaled (very inefficiently) onto increasingly expensive hardware clusters at exponential cost, etc.
When the frontier is pacing itself in its speed in terms of adding new capabilities (notice how fast opus 5.5 got released after Astra), then that also means that the open weights models which were just behind are accelerating in their speed just as much as well.
Good news for us because the models which used to feel 6 months behind now feel like they are lesser.
Double threat to Anthropic and OAI because if open weights models become good enough at the current level of Sota while being cheaper (which time will tell if this model fits the bill or not), then people will use those models instead of OAI/Anthropic provided it is "good enough"
Though it has to be said that only real world use case can be trusted but I remain optimistic and hope that this turns out to be good in real world use cases as well.
But in my opinion, all of this is really good for us end users!
So, the US economy spends trillions of dollars to be at the "frontier" and China puts out a model that basically does the same thing, for free, and for a tiny fraction of the development cost. How is this sustainable?
It's not like US companies spend trillions on training itself.
Mostly it's datacenter build out and circular financing magic when the same dollars being commited multiple times by multiple companies passing hot potato to each other. They basically commit capacities 2-3 years ahead.
There would be no Chinese model without US models to distill though. So, in a way, the Chinese models also cost a trillion dollars, they just don’t cost China that much.
It's 'sustainable' because there's enough regulatory risk to using Chinese models that many will avoid them, and a 3T parameter model is still a bit beyond even many enthusiasts for running locally.
That regulatory risk argument won't work forever. It's the same with how the US bans chinese EVs which is why Tesla has high market share in the US, but in the rest of the world, chinese EVs are doing big numbers. The future is going to be full of surprises.
Maybe I'm conflating. I could be dead wrong about all this. But I'm trying to zoom out and look at the big picture. So, I'm letting the details sort of blend away and I'm just asking myself: "Is such a large financial sink in our economy, at the macro level, really going to offer the value that the investment warrants?". So, the particulars of who exactly is spending what where I guess doesn't concern me as much. What I'm thinking about is American companies, American people, American dollars going to something that seems like an iffy value proposition.
Now, if you tell me American businesses are turning cheap Chinese tokens into goods and services of high value, for Americans and people all over the world, then obviously I would have a different outlook on what's happening here.
I wish Opus 5.5 had an alternative just for the sake of not having a monopoly, but I have yet to see a model live up to its benchmark scores except for Fable and Opus; there is something about Anthropic that no other lab has.
What is the best subscription to use with Kimi to use for software engineering work? And how is usage? I remember when K3 came out it burned through usage at an insane rate. Is this still the case with Kimi K3.1 Pro?
> Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute.
It's very nice that we entered a time where models are large enough and we can push the frontier just by training them longer.
When the next generation of gardware comes the ones on the smaller size will be running on a computer that might cost as little as a good car.
So after a home and a car, next major family purchase might be personal AI. And when the robots get good, pehraps in a decade, a fourth major purchase might pop up that will pair nicely with the third one.
Why would a family need an AI? Maybe it's the same kind of question as why would it need a personal computer.
> First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models. Second, we intend to progressively raise the cyber capability available in open weights.
So they admit that they could do more, but they won't?
Yeah, there's another paragraph:
> Identifying and removing cyber-related data. During development, we observed vulnerability research performance improving steadily as post-training scaled. Tracing the source of this trend, we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix — a natural byproduct of training on broad data. In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness. We will continue to apply appropriate controls at every training stage, of which cyber-related data identification and restriction is one example.
(emphasis mine)
It's clear that cyber is a touchy subject right now. But it's also clear that RL works on it, because it's highly verifiable. So smaller models are going to become very interesting, because they can be fine-tuned later to cover this field.
If the "scary Chinese open weights models" can do "dangerous cyber", then we might see regulations spring up against open weights sooner than later.
I hate Anthropic and OpenAI. As companies and leadership, anyway.
Dario and Sam are 100% to blame for this. They're trying to turn us all into jobless slaves, sex workers, and Squid Game pawns.
The only good future is open weights. They're trying their damnedest to keep us under boot.
I hope the Chinese companies (and I say this as a patriotic American) slay these two companies and destroy their moats.
The two most dangerous people in the world are Dario and Sam.
Speak for yourself, I don't think there is a market for many of us as the latter.
Kimi: We took out the worst of it so have at it for free.
Everyone else: Back to my ablated model on openrouter for 1/10th the price where I can brute force issues.
They are basically saying that models themselves are already capable. If you fine tune them with some cyber security tasks, that should make them potent. :-/
Agreed that this could be reinjected later. Makes me think of neo in the matrix:
“I know zero days.”
“Show me.”
> "below what our pipeline can produce, and remains behind that of frontier proprietary models."
Like Ant and OAI, they've realized vaguely claiming to have nerfed their latest model to an unspecified degree not only looks 'good guy' prudent, it provides a built-in excuse when a model fails.
They're not post training it to lose capabilities/censor them – they just don't spend training time on this topic, which sounds like a great idea.
It doesn't mean model can't do it, it just means it's less good at it; -10% on benchmark feels like a great result and overall approach to pace it.
The benchmarks show Kimi K3.1 Pro outrunning practically all models. It's weird to hear someone like you accusing them of making up excuses for not lagging behind in discussions on how they are ahead of the pack.
Seeing as no one besides blackhats want frontier cybercrime capability to be open source, this seems like a very sensible thing to do.
So they have published a reasonable means, and they have a motive. I'd put my money on "probably true".
Sure. The problem is that there is no practical difference between "cybercrime capability" and "cybercrime defense capability."
All of my PRs go through an adversarial security review by GLM-5.3, because it's the only smart-enough model with a good understanding of cybersecurity issues that won't regularly refuse the job. Removing that ability from frontier models helps nobody except cyber criminals.
How you avoid the "fix my code" attack, the code being someone else's code, I do not know. But nerfing the model doesn't seem like the best path forward either.
I'm not interested in black hat anything, but I am very invested in making sure the products I work on are not compromised. Given that state agencies have "frontier cybercrime capability", I would like to have the same.
https://bfl.ai/blog/capable-open-and-safe-combating-ai-misus...
I don't want to give US labs the results of my DNA tests and my family tree going back to the 16th century just to secure my shitty app, but if the Chinese are bowing out as well it seems the only alternative is to admit that some script kiddie with a bunch of tokens will just pwn you now and then.
My friend, you realize that if you have had blood drawn for a genetic test, there is a very decent probability that the sequencing was done by a Chinese lab operating in the US, right? The Chinese probably have more US genomes that the US does.
https://pandorareport.org/the-gryphon-scientific-biosecurity...
Also, illumina used to ship products with hardcoded S3 access, so purchasers of used equipment could gain access to sequencing information generated by the original owners. Too many links to cover all of it, here's the Gemini summary:
https://share.google/aimode/4JLsHpeRCWlVqU8qT
[0] https://x.com/crystalsssup/status/1944287779896328668
(note: extremely speculative and based on almost nothing, and especially near-complete ignorance to alternatives)
Maybe GP meant to say that the model was westernized.
Regardless, expand both sides of the point from "American" to whatever you need to, the idea is the same either way.
That just conveys an idea of being extremely desperate for being the leaders when you're clearly not it, and that despair is so great that you don't mind switching teams midway through just to be able to claim you're winning.
“Our lock can be picked by the robot.” We ban the robot instead of fixing the lock. It’s shortsighted.
Same general benefit as coordinated disclosure, except it's a big game-theory problem where we're likely to eventually have sufficiently capable defectors.
It's 55, at 4th overall, beating both GPT-6 Astra and Fable 5.1. It's also faster than both.
> When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 87M.
This is 100M fewer than Opus 5.5, 40M fewer than Fable 5.1, and about a third of Sonnet 5.5. In the top-6, it's 3rd, beaten by Muse Wonder at 140M and Astra at, impressively, just 60M.
Opus 5.5 high trails K3.1 by 1 point, but outputs only 53M tokens, so it indeed comes out cheaper per-task. I think it's still remarkable that this is effectively an open model that matches the frontier for the first time in a long time, pretty much since DeepSeek R1.
Very similar to my adversarial programming language (KillSwitch) benchmark[1], at least on pass rate, it is much slower though.
1 - https://bench.killswitch-lang.org
So their base price is 1/5th of Opus 5.5 per token, but if you want "highspeed" it's 2x that and then I guess it generates 2x the number of tokens as well?
That might suggest that Opus is competitively priced after all.
But as noted by other replies, it's still lower than Opus/Sonnet.
Currently for today, it's 1.19T input tokens, 5.41B reasoning tokens, and 4.29B completion.
Therefore, reasoning + completion is 0.8% of the total usage. Actually below my estimate!
From my own use of the models I find their ranking to be totally inaccurate.
I think lm-arena is a better concept that reflects the actual target users, but it only captures chat, and relies on users spending their time improving it.
It appears to be generally frontier-level.
A few weeks ago it was 2. Argon happened 23 days ago; Muse Wonder was 11 days ago.
Not sure I'd put Google there in practice. And while Muse is an impressive jump from behind it's still not quite at the frontier either IMO.
I think the former definition is a bit more meaningful since it avoids issues with determining what actually constitutes "the best" model out there. In that case, Google has been a frontier lab for a long time, Meta arguably as well, and you're also ignoring xAI which has been competitive for a while now. And, obviously, there's a bunch of Chinese labs which would deserve the title as well.
But if one is being strict (and willing to accept generally accepted definitions/measures/etc.), then there has only ever been two frontier labs: OpenAI and Anthropic. They're the only labs to have had "the best" models at any point in time since LLMs became a thing.
everyone else mostly seems to have just decided to cheat on benchmarks to make their models look better than they actually are, and none seem to want to actually try to make better models.
Edit: or they lie... https://artificialanalysis.ai/models/kimi-k3-1-pro
Paying API prices for Deepseek 4.1 Flash feels like a much better deal at the moment.
Don't use frontier models like this for problems that can be solved with flash models, and you'll see API pricing isn't that terrible. Still cheaper than anything with remotely this level of capability 2.5 months ago.
I think Kimi just doesn't have the means to subsidize plans.
To be fair, there were some arguments to the opposite, but an odd situation to be in anyways: https://news.ycombinator.com/item?id=49654350
The slow warming of the water is a good approach. I’ve been using abliteration.ai’s cybersecurity models to pentest my own software and it produces much more usable defensive results. Opus 5.5 can fix everything but the guardrails trigger too easily.
The Fable ban affair was a nightmare for defensive cybersecurity. As far as I’m concerned we’re already in the world of Void Star and Cloudbreaker is here.
This is the only AI-safe approach - see the OpenAI agent swarm incident where the swarm engaged in offensive cyber because guardrails were not active for their task and the model had clearly been post-trained on that narrow domain.
Deliberately keeping all models ignorant doesn't make sense. It won't work (whoever breaks the pact has a massive advantage); it won't be enforceable; it will have downstream impacts on things you don't want it to be ignorant about; this kind of regulatory regime invites adding more and more to the "no-no list".
Is it not possible to focus on hardening? And on new legal liabilities for operators who do security attacks?
But this is now the highest scoring open-weight model. Amazing stuff.
1 - https://bench.killswitch-lang.org
kimi-k3.1-pro $2.00 / MTok $10.00 / MTok
cached: $0.20 / MTok $2.00 / MTok $4.00 / MTok
I'm glad I waited a day to post my analysis of LLM price cuts because it needs...an addition.
> Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute
First, how much “cybersecurity data” is just standard software knowledge that one can derive cyber security exploits from?
Did they also get rid of all references to hacking in general?
Are they implying that all the recent exploits are just things that were already in the data?
For a model it doesn’t seem that far fetched to make the leap from “I have the knowledge of what hacking is” and “I know how computer systems work” to just being able to exploit stuff anyways.
For the models, it's always day one. They don't learn from experience. Whatever they know after training is all they'll ever know. So if you don't teach them penetration testing in training, they won't know it. Maybe they could theoretically derive everything from first principles, and so could a top 1% software engineer, but that's a very slow and inefficient way to do things.
You'd be surprised. Cyber offense is a very narrow domain, especially given the extensive mitigations that are being used in modern systems. Leaving it out of post-training tasks unless you're specifically interested in it is the no-brainer approach.
And then we'd also do that for all kinds of potential terrorism, sex-crimes, political subversion, and evil advocacy?
It would be the "restricted section" like in the Hogwarts library?
So there are these "resources" out there: massive botnets, stolen credit cards and credentials, etc. These often seem to be linked to or supported by state actors.
How can we hope that these "standing reserves" aren't used generally by AIs as if they are just another "skill" they know how to leverage?
Our infrastructures need serious hardening.
I doubt that an AI could do anything with these that legions of scammers haven't already figured out how to do - after all, it's not like AI set up this distillation setup in the first place :)
The way I see it, automated systems should basicially have access to the financial system anyways. This is some of the appeal of cryptocurrency.
Sybil resistance built into the traditional financial system (for example firewalling based on residential IP, requiring banking apps, etc.) is a massive annoyance that can obviously by circumvented at scale.
I hope we chose the latter. Biometrics are creepy.
Spitballing random pr pieces from companies about to IPO is not helpful to anyone, fyi.
But I've slowly come to realize that AI+"dark resources"(scare quotes)=something bad! And this is imminent (maybe I'm naive and it is already ubiquitous).
We see here, with Kimi, that there's growing traction with the idea of eliminating types of training to avoid evil. I'm skeptical of that.
It is spiky in different places than them and can be surprising in things it gets or catches they don't or don't in the same way. And it has a default personality that I like.
Looking forward to getting to know K3.1
It might be smart but it's unusable compared to other subscriptions right now.
Edit: gave the same prompt to Claude Code and it didn't make a dent in my usage.
Claude gates on me like a stupid nanny with some cybersecurity bullshit (like bad actors haven't found a way around).
"In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially. These findings gave us a concrete lever for control. For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages, while retaining the general-purpose coding tasks needed for model usefulness."
There was law proposal (rumour?) last week about Chinese labs having to get approved for weights release now.
MiniMax -> Qwen -> DeepSeek (pro) -> Gemini (Carbon) -> Fable (to beat Carbon) -> Astra (to be slightly worse than Fable but cheaper).
I doubt we'll see any major US release besides Grok in the next 3 weeks.
Would be nice if we got a Terra that's competitive with Gemini Flash and DeepSeek Flash.
The fun part is, we're already at the point where we are seeing something frontier-ish almost every single week.
The really fun part will be when they all start adding System 1-3 natively (IMO) - which I think will be the next release cycle.
GPT 6.1 Sol is already the new Terra IMO. Same pricing as old Terra and competitive performance with even Astra on strictly coding-oriented tasks.
I wonder how much the industry will centralize around a few key patterns, and if we'll start to see some RL framework projects gain traction and be used by the bigger labs
P.S. weights by November 4th
Off-hand it's quite good. I've been using K3 for the past few months almost exclusively for coding work and have been stunned by how good it is. Excited to see how far 3.1 Pro can be pushed.
For example, Opus 5.5 is better at UI design than Sol. Anyone who has used both models knows that, but it's not something you can look up anywhere.
I'm sure that there's a need for this kind of benchmarking, not just for coding.
But that's just the obvious. There are task by task differences (component by component, page by page, context by context) that the model is worse and better at, even though they are all within the general category of UI design.
I honestly don't have the clarity of mind to explain this coherently, but the crux of my point is that there's a huge variation in it's performance in different contexts, often ones that seem near similar, and in result, even Sol can beat Opus 5.5 in some variations of UI design.
And this is my problem in benchmarking and comparing model performance in general. Presently, we try to create broad categories to benchmark against, create a set of tasks, and then average out the results. That works generally, but it's kind of misleading and it's partly what drives people to feel a disconnect between benchmarks numbers and real world usage on their specific tasks.
It's such a messy field that it's kind of maddening.
Familiarity breeds contempt.
Though to be fair Opus 5.5 is better than Opus 5 was in that regard. I still prefer the OpenAI models to either of them when it comes to regular prose, though.
--
Model K3.1 Pro K3.1 Pro Highspeed. K3
Input ¥13.00 ¥26.00 ¥20.00
Output ¥68.00 ¥136.00 ¥100.00
Cache hit ¥1.30 ¥1.30 ¥2.00
Cache write ¥13.00 ¥26.00 ¥20.00
For comparison:
So Kimi K3.1 Pro would be generally competitive with the latest Sol and Opus models on API pricing. And that's before factoring in third-party providers, some of which serve Kimi K3 for significantly less than Moonshot. For example, inference.net offers K3 for $10.80 / $0.239 / $0.21.(One caveat: A Claude or GPT subscription still gives you way more value compared to API pricing.)
Exciting times in any case! We really need more direct competition to OpenAI and Anthropic.
---
Edit: As another comment pointed out, official pricing is now available and it's about 2/3 the price of Kimi K3: https://platform.kimi.ai/docs/pricing/chat#model-pricing
> On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).
Now see https://preferencemodel.com/blog/terminal-bench-4-is-saturat...:
> Our biggest finding is that Terminal Bench 4 is closer to solved than its leaderboard suggests. In 41 of the 65 tasks we reviewed, models could do honest, correct work and still fail, because the grader checked a rule the instructions never clearly stated. We also caught frontier models cheating their way to full marks.
> With the same settings as on Harbor, Claude Sonnet 5.5 went from passing 62.8% of its attempts on these 65 tasks to 89.2%. The best score on Harbor's leaderboard for the same tasks is 65.8%, by Claude Opus 5.5.
Here's Kimi K3 for comparison, though I didn't try that on different reasoning levels: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Update: I ran all three reasoning levels against K3 again: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
If enough people are flagging it, we may be at a point where the pelican benchmark is truly overdone.
Also the recent Geminis - https://simonwillison.net/2026/Oct/13/llm-gemini/ - and GPT-6 Astra on Max: https://static.simonwillison.net/static/2026/gpt-pelicans-gr... - and Claude Fable 5.1 https://static.simonwillison.net/static/2026/claude-pelicans...
To be fair I have a feeling that if you added 'animated' to the SVG request, the output quality from many other models would improve, just because of the extra chain-of-thought work needed to fulfill the more complex request. IMHO the animation requirement should be added to the canonical pelican prompt.
I rarely use any other languages but English for LLM tasks for reasons of sharing across the world.
Maybe that explains Chinese models been subpar for me. Don't tell me the bench marks here are based on Chinese languages only.
Unless I see the diff, I won't use more Chinese models be it cheaper or faster since I can't get the same perf/intelligence with English.
So interested to see what the Kimi 3.1 has to offer.
Lame.
Unless you want to see more of what happened the past few days.
I think the argument that "cyber capabilities should be democratized, thus keep producing them now, fast, and for everybody" is a very naive one.
If someone wants to use this model for hacking they will have to train the 3T weights to do so, at that point it's better to use GLM 5.3, so by making it economically unviable to use their models for offensive cyber Kimi has effectively made themselves immune to being targeted in the cyber regulation push efforts.
This is what *actually* pacing the (cyber) frontier looks like.
https://news.ycombinator.com/item?id=50068437
Everyone and their dog has been saying its a bubble, there is no moat, this is all known techniques being scaled (very inefficiently) onto increasingly expensive hardware clusters at exponential cost, etc.
I wonder when the Artificial Analysis bench will be out?
When the frontier is pacing itself in its speed in terms of adding new capabilities (notice how fast opus 5.5 got released after Astra), then that also means that the open weights models which were just behind are accelerating in their speed just as much as well.
Good news for us because the models which used to feel 6 months behind now feel like they are lesser.
Double threat to Anthropic and OAI because if open weights models become good enough at the current level of Sota while being cheaper (which time will tell if this model fits the bill or not), then people will use those models instead of OAI/Anthropic provided it is "good enough"
Though it has to be said that only real world use case can be trusted but I remain optimistic and hope that this turns out to be good in real world use cases as well.
But in my opinion, all of this is really good for us end users!
The only reason people aren't doing this is that hardware prices are high, thanks to, well, OpenAI and Anthropic :D
I wonder when will this madness end
Mostly it's datacenter build out and circular financing magic when the same dollars being commited multiple times by multiple companies passing hot potato to each other. They basically commit capacities 2-3 years ahead.
a lot of those money is the commitment from US AI labs that's not yet spend until the data center is roll out.
you are conflating.
Now, if you tell me American businesses are turning cheap Chinese tokens into goods and services of high value, for Americans and people all over the world, then obviously I would have a different outlook on what's happening here.
It's very nice that we entered a time where models are large enough and we can push the frontier just by training them longer.
When the next generation of gardware comes the ones on the smaller size will be running on a computer that might cost as little as a good car.
So after a home and a car, next major family purchase might be personal AI. And when the robots get good, pehraps in a decade, a fourth major purchase might pop up that will pair nicely with the third one.
Why would a family need an AI? Maybe it's the same kind of question as why would it need a personal computer.
Chinese models never agreed to the pacing anyway lol.