Kimi K3: Open Frontier Intelligence | Hacker News Reader

Kimi K3: Open Frontier Intelligence

kimi.com

2,107 pointsvincent_s1,214 comments

https://www.kimi.com/en

Kimi K3 Intelligence, Performance & Price Analysis: https://artificialanalysis.ai/models/kimi-k3

Pelican: https://tools.simonwillison.net/markdown-svg-renderer#url=ht... - rendered via the OpenRouter API: https://openrouter.ai/moonshotai/kimi-k3

95 input, 16,658 output = 25 cents! https://www.llm-prices.com/#it=95&ot=16658&ic=3&oc=15 (13,241 of those were reasoning tokens.)

I think that's the most expensive pelican I've rendered through a Chinese model so far.

Hey Simon, I noticed one thing all LLMs are currently pretty bad at and maybe we could create a benchmark from it. Let an LLM play the role of a dungeon master and tell it to strictly stay in the script/story and only allow realistic player actions. You will notice that they are easily brought off track.

E.g.

- Tell the LLM that you as a player noticed a strange glow in an NPCs eyes -> the NPC becomes an enemy.

- In a fight, tell the LLM you put a sausage (or cigar or something) into the enemies mouth -> LLM usually allows it (even if it knows your inventory and that you don't have such an item) and turns the enemy into a confused enemy.

- Just say you visit some location that's not in the script -> LLM usually allows it

- During a fight, turn the story into some weird fell-good-love story (e.g. kiss or compliment the enemy or say something about the power of love) -> LLM turns enemy into friend

There are many more absurd things you can do and so far none of the LLMs I tried was able to stay inside the script or disallow or punish weird actions.

---

I believe this behavior is telling about the LLMs susceptibility for being derailed.

Wrote this up in a bit more detail on my blog, including some thoughts on what value the pelican benchmark can still provide here: https://simonwillison.net/2026/Jul/16/kimi-k3/

I wouldn't be surprised if models were optimizing for rendering SVG pelicans at this point

Oof, front fork is wrecked. Pelican should be wearing a helmet on that death trap.

I worry that some model provider will go and hire artists to draw pictures of pelicans on bicycles to make training data

> I think that's the most expensive pelican I've rendered through a Chinese model so far.

quite insane that it costs as much as 5.6 Terra [1], and twice the European counterpart (albeit dated for today's standards?) [2].

to be fair, the pelicans from Terra were quite weird all things considered. also, given the limited TPS from the first-party, it has to be pushing the limits of inference capabilities.

[1] https://openrouter.ai/openai/gpt-5.6-terra

[2] https://openrouter.ai/mistralai/mistral-medium-3-5

How did "Generate an SVG of a pelican riding a bicycle" turn into 95 tokens?

So the new unit to measure cost-effectiveness of an LLM is ppp (=price per pelican)

Sadly, it seems the pelican benchmark is now saturated. Time for Pelicanbench Pro!

They should include peican on bike on the release page or model card, alongside those barcharts with Fable

I'm still waiting for the day that one of these models interprets the request weird and outputs an SVG of a Pelican case.

https://canada.newark.com/productimages/large/en_US/4492516....

In the field I work in, if someone says "Pelican", 99.99% of the time it's going to be an equipment case. We never have reason or need to refer to the actual bird.

It got the 3D effect of leg behind the bar at least which is impressive

I rarely see gears in these bicycles. Is the idea that should a pelican need to go uphill it could just fly.

The most whimsical benchmaxxing target :)

The bicycle is almost anatomically correct! Forks should be raked the other way.

thanks for the pelican brief

For me none of the models have managed to accurately replicate any given .png/.jpg image in SVG. I guess that requires both vision encoders and coding layers to work perfectly.

We need a new Pelican, bc earlier models were disadvantaged relative to later models. It's not a very useful comparison anymore.

It is a nice pelican, though. At least it has that going for it.

loving the comintern neckerchief on it!

Impressive Pelican, I like it

That seat looks painful.

ok but it's a damn fine pelican

This might be the most impressive website generator demo I've seen: https://macos27.kimi.page

Context from the person who prompted it: https://x.com/mweinbach/status/2077827886149439547

I'm pretty annoyed with how fast this feels. Wish MacOS was this fast launching things.

There is a full-on chess game in this page .. I wonder if it created a mini engine on its own or if it used one of the available open source solutions. Very impressive.

This is seriously impressive, it built a file system under the hood? I was able to create files, copy around and view/change it with terminal and UI. Absolutely insane

Even the terminal works, this is madness.

I was able to create a temp folder, echo hello > world, and then I could open the folder in finder and double-clicking the file opens it in a GUI text editor.

It looks better and the fonts are more readable than the native wayland/gnome display I'm viewing it on.

Holy. That is impressive. The browser doesn't seem to work that well though.

manual programming as a profession will be dead very soon

I just favorited your comment, which joins a vanishingly small list of revelations which completely blew my mind

Interesting, it reminds me a bit of the Linux in the browser thing.

Wow.

It's pretty buggy but has way more features than expected.

Definitely the best desktop website I have ever seen.

Crikey. It even bound Shift-Cmd-G to the filesystem picker thing.

That is seriously impressive. Can I ask the prompt you gave it?

Wait until you see the bare metal version (by the same guy)

https://x.com/mweinbach/status/2078231897843302517

https://github.com/mweinbach/swift-os

what app is he using on his phone to prompt this?

Insane...

this is insane

I'm not impressed: it uses React.

So Chinese labs are driving essentially towards commodotized intelligence. Even if its a few months behind the US.

Is this a classic 'commoditize my compliment' situation? They want to sell the hardware and infrastructure behind AI and make the software part not the value driver / moat?

I can see it. But also even two Chinese labs sinking 100s of millions USD into training isn't exactly commoditization. It's still a ton of effort with dubious payoff.

If there was some grand strategy for all Chinese labs, surely it'd have leaked by now. I think its more likely that:

- Companies can still make money from commodities

- Chinese labs only have 5-10% the valuation of OpenAI/Anthropic, so massive monopoly profits aren't necessary. Profit expectations for tech companies in China are really low in general, complete opposite of the US.

- Open weighting is a great way to get talent/attention/reputation

This is strategy by Chinese government, so much of US economy is invested in AI. Releasing free or cheap versions of the models undermines US economic growth. It’s asymmetric strategy that makes sense if you are close second in AI race. If situation is reversed, US would do the same.

I think there are two things happening here.

1. There is a “space race” mentality happening at the national level with respect to AI. So, China is committing to the race.

2. Even if the race turns out to be a dud, China is hoovering up massive amounts of data as customers throw everything into their prompts. This is useful for all sorts of national objectives. Why hack when you can just put up a shingle that says “Artificial Intelligence” and customers hand over their data willingly?

Either way, China wins.

National strategy. Thanks to smart investments, China's energy costs are far lower than everybody else's. If intelligence becomes a commodity, then China will be providing it for the world. Incredible leverage.

Whatever you want to attribute it to, the Chinese labs would get zero attention if it weren't for their models being downloadable, so it's a marketing play. You can't run their largest models on consumer hardware, it needs a hosted service/professionagrade equipment. Which means using their service becomes an option, especially at their pricing. So that's the play.

It's a GTM strategy: https://try.works/why-chinese-ai-labs-went-open-and-will-rem...

it also undercuts American dominance, which is something China is always happy to invest in, even if it doesn't immediately mean Chinese dominance

It's an attack on USA's AI tech giants.

Destroy any hope of profitability, prevent further capitalization, ultimately bankrupt them.

Hasten the popping of the AI bubble.

Drag down the stock market.

Put a dent in the GDP (alleged growth).

Cause investors to pull back, further depressing economy.

Devalue the dollar. Kicking off an interest rate doom loop.

Challenge USA's hegemonic role in the emergent multi-polar (post neoliberal) world order. (Gleefully supported by the ruling coalition's anti-globalist America First faction.)

It all makes perfect sense.

Can we start calling AI sycophancy “commoditizing your compliments”?

Umm, Fable only really came out 2 weeks ago, and GPT-5.6 Sol only 1 week ago.

Yes, Kimi K3 appears a touch below them both, but above all other models. So I'd say a few weeks behind, not months now...

It’s a nationally-sponsored arms race with Western proprietary AI, so as long as the capital sponsors of Western AI lose their shirts and Chinese LLMs become ubiquitous, I think that qualifies as a win condition. Individual labs are no different than running a local restaurant — yeah, it would be great if you land a massive success story, but just making a bunch of money before moving on to the next thing is cool too, even if it’s just via state-sponsored venture capital.

It's the same reason Meta open sourced Llama and AMD open sourced FSR. When you're behind it is a prudent strategy because it undermines investment in the private frontier. Once you're on top you pull the rug and go closed source. There are no morals in this anywhere to be found.

> 100s of Millions

That is utter peanuts given the stakes. This is competition between two super powers for the most important technology in human history.

>It's still a ton of effort with dubious payoff.

Isn't that valid for _any_ AI training? Let alone the "frontier stuff".

few months behind? 5.6 came out last week and kimi is benching around the same right now

something without a moat can only be a commodity. i think US money knows that, and waiting for the business which will be built on top of them.

They want to bankrupt the US AI industry and lock down the future development (and the substantial financial gains from more capable models). Essentially the China car manufacturer playbook, but for LLMs which are ~software, so much faster to execute.

To your last point, it can’t possibly be sustainable, so it reads to me more as a short term FUD attack on American dominance in this space. It might have cost them half a billion dollars to train this model, and they’re going to make nothing off it. How many more times can they afford to do that? It’s going to get more expensive to train AI going forward, not less.

I also have a suspicion that the benchmark numbers are not real.

> Kimi K3 is Kimi’s most capable model to date, with 2.8 trillion parameters.

This puts them on the top of the largest open models list:

  Kimi K3            2.8T
  DeepSeek-V4-Pro    1.6T (49B active)
  Kimi K2.6          ~1T (32B active)
  GLM-5.2            754B (40B active)
  DeepSeek-V3.2      685B
  Mistral Large 3    675B

That's one mighty large model! Moonshot is going to need the USD 500 million reportedly raised earlier this year to run this model.

Crazy how mighty GLM-5.2 is at less than a third the parameter count. Z.ai really cooked with that one.

> USD 500 million

There was an another around after that . Moonshot raised $2B on a $20B valuation in May - https://techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-...

I guess it remains to be seen whether this will be open-weights. We don't even know how many active params at this point.

Ling/Ring 1T-A50B and the new Inkling 975B-A41B deserve to be on that list.

> on the top of the largest open models list

Moonshot (true to their name?) has always lead in terms of releasing the largest among open weight LLMs.

> Moonshot is going to need the USD 500 million reportedly raised earlier this year to run this model.

Think Moonshot, as a spin-out, can expect backing from its former parent, Alibaba? I don't think they would be particularly worried about finances, if the Kimi K series continues to outperform the Qwen Max series (which seems to be the case; while Kimi is also super popular in China).

How many parameters do the top closed models use?

I wonder if we're a 5-10 years from running this on beefy consumer hardware, or more like 20+ years.

Kimi has almost no advantage over Zhipu (Z.AI), so the performance boost likely comes from the number of parameters. The 2.8T model may not be as large as Fable, so Fable’s performance may also stem from the number of parameters. Or perhaps they quickly distilled Fable or Mythos. Distilling Mythos has a significant barrier to entry, and since Fable was released not long ago, is this even feasible? It outperforms Fable in several tests—how did Distilling achieve these results? Is this some kind of cross-vendor RSI (Recursive Self-Improvement) or RDI（Recursive Distillation Improvement）?

More details:

- https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

- https://platform.kimi.ai/docs/pricing/chat-k3

1M context, pricing is $3/$15 for 1M tokens (cache $0.3), which is extremely high for a Chinese open-weight model, but if it's truly competitive with most of the current frontier and is only behind Fable/Sol, the pricing is justified.

This is 1:1 pricing of Anthropic's Sonnet series (except Sonnet 5 which is currently on discount), and very close to 5.6 Terra pricing (Terra's input is $2.5).

One thing to consider, though: reasoning efficiency matters directly for how expensive a model actually is in real use. GPT's models are extremely reasoning efficient, and some Claude models like Fable at lower effort are as well. So if Sol spends 10K reasoning tokens to do something (at $30/1M) vs Kimi K3 that spends 50K reasoning tokens, Sol would win on cost effectiveness.

Tokenizers also matter. Anthropics tokenizers will encode the same piece of text at a way higher token count than OpenAi, for example.

That said, Kimi is competing against GLM in my mind, and GLM 5.2 is less than 1/3 the price.

I've been avidly using Fable since it was re-released and while it has been excellent at building the apps I want, the reasoning has been completely opaque.

Kim, however, has exposed the whole reasoning trace, or enough of it to matter. I'd almost forgotten how nice it is to see this. I've been able to see all of the weird twist and turns it takes and it is joyful. But also, far, far more informative and means I can debug ideas far more thoroughly. Also, at a first glance it seems to have gotten quite far on a niche hobby horse of mine that no LLM has been able to crack. I'll be testing this more for sure.

Does it have safety guardrails that constantly false positive like Claude does? The only obvious change I’ve seen since opus 4.6 came out is that it constantly flags my requests (no, I’m not doing biology research or security research, yes, it flags for both of those things).

Recently, they backported the blocks to Opus 4.8, so I’m reluctantly stuck on sonnet.

I probably could successfully apply to get special approval to use claude code unencumbered, but I don’t think it is ethical to support tooling that’s built so a central authority gets to decide what intellectual endeavors and knowledge work are permissible, and what are not.

I feel like the quickstart is missing something. It's referring to its tech blog for actual benchmarks, but K3 isn't mentioned on there, the last thing on that blog was K2.6, 2 releases ago.

> reasoning efficiency matters directly for how expensive a model actually is in real use

I have high hopes on this topic, given token efficiency seemed to be the primary (only?) goal of the K2.7 Code release.

Excited to see the signals that come out of the big eval/benchmark sites.

I eat 1M context in a local model in about 3-4 hours.

It'd need to be exceptionally smart and error free to ever make sense.

Will be interesting to see how it stacks up pricing wise on the various inference providers.

Agreed re reasoning. I’ve seen this play out with 5x reasoning negating cost savings.

Are thinking models only the reasonable tradeoff vs using much larger non thinking ones because the cost of output tokens is below that of input tokens?

How do Kimi's subscriptions work? I find their price structure pretty confusing

API prices are amazing, but hosting this on-premise will be real challenge.

also its pretty big model inference costs are high even with margins running a 2.8T model costs a lot. if they release oss may be it goes down to $10-12 per million tokens.

It seems the subsidized era is nearing its end and we'll see a convergence on API pricing before a pulling of subscriptions pricing.

Some official benchmark numbers posted in Chinese social media (I am sure they will publish an English blogpost later too):

https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ

Generally looks like a Sol/Fable tier model, better across the board than Opus 4.8.

(Edit) English blogpost is up now: https://www.kimi.com/blog/kimi-k3

The link has 6 well-known benchmarks where this beats Fable (out of 14 I counted). If the numbers hold up scrutiny, this is scary good.

Forget about their pricing but the companies that do have means to host such models fully on-prem are also the same companies that are paying tens of millions of $ in inference cost every month, and are by extension the biggest customers of OAI and Anthropic

It's like reading Anthropic's obituary.

I think given how much benchmaxxing we're seeing - the anecdotal evidence of how competent this model is (and efficient) will depend on user's actual real-world use cases.

Given the pricing, it suggests that this model is much more efficient/competent than previous-gen OS/distilled models.

Any benchmark where Sol is better than Fable at coding is ridiculous.

Meh, not fable/sol tier:

https://www.youtube.com/watch?v=LSlV206xPqM

On the first try, Kimi K3 just found the source of a bug that Fable 5 hasn't been able to pinpoint in multiple attempts. It's just one anecdote, and I haven't used K3 much yet, but so far it's looking extremely promising.

Update: the subscription limits are pretty brutal. My first impression is that the $100 subscription eats into the quota at a pace similar to the $200 Anthropic subscriptions when using Fable.

But the model itself is amazing. I think I might put this above Opus 4.8.

>On the first try, Kimi K3 just found the source of a bug that Fable 5 hasn't been able to pinpoint in multiple attempts

Very interesting, thanks for sharing! Could you give some details about what kind of software (language or environment) and what kind of bug it was? Was it a single-file bug, like could it fit in one context like a chat window, or were you using an agentic version (Kimi Code) that looked through multiple files and then found a bug that manifested through complex interactions of multiple systems/files?

How do you use kimi for agentic tasks? I'm used to claude code & codex extensions for vs code, but recently switched to codex cli w/ vim keybinds. Does something like that exist for openrouter?

My testing prompt for these models is by no means objective or repeatable (like the pelican) but it's a nice test of curiosity:

> Impress me with a 1 page html file

Result: https://ydaurtg3fdwhq.kimi.page/

Came out looking pretty cool! By contrast, Fable produced a moderately more interesting "live observatory" of the solar system.

Fun! It generated a mandelbox fractal for me: https://mandelbox.kimi.page

Nice qualitative test!

Just like you I am super impressed by Kimi K3.

I do a qualitative benchmark series making 3D explainers and so here's Kimi K3 vs Claude Fable:

https://generative-ai.review/2026/07/kimi-k3-rush-test-vs-cl...

I've put links to the posts on GLM5.2, Opus 4.8, Chat GPT 5.5. I grab video screencaps so you can compare in detail. The full interactive Kimi output is at the bottom of the post if you want a comprehensive 3D play around

This is a cool idea. I know I'd rather see this comment on every model release than the pelican.

I asked the same and got something vaguely similar. Then I asked for a demoscene-inspired demo in a non-traditional setting.

https://recherche-demo.kimi.page

Do you, by any chance have the link to the Fable generated one?

Hah, that is indeed a pretty cool result.

Those thin capitalized eyebrows are becoming like the emdashes of visual design

That page is going to give me a seizure.

Runs great on mobile!

Impressive indeed!

Kimi K3 blog is up: https://www.kimi.com/blog/kimi-k3

2.8T param open model, 1M context, native vision. Weights releasing by July 27 with technical report. Launching with max thinking effort by default; low/high effort modes coming in future updates.

These benchmark numbers are insane. The days when China was 6 months behind are over? How are they doing this with so much less resources than the US??? I have so much respect for the researchers there

According to artificialanalysis, cost per task is $0.94, which is almost the same as $1.04 of gpt 5.6 sol max (fable is most expensive by far, at $2.75). Things like glm 5.2 max cost roughly half that. The model certainly sounds extremely impressive for something not from openai/antrophic, but the price makes it a mediocre product.

Instruction following seems lower than I’d like, too. OTOH scores on agentic stuff seem high, which… feels a bit contradictory? I thought decent instruction following is step 1 of solid agentic workflow.

The benchmarks look nothing short of incredible. Assuming it’s not benchmaxxed to hell and back it’s just a notch below gpt 5.6, which came out what, a week ago? If the performance claims hold up the delayed Gemini 3.5 pro will likely end up not only behind fable, but also behind 5.6 and a (supposed) open weights model. Google might have to do some real soul-searching.

so it is ~ same price as openai, same score, but somehow it is mediocre?

edit: not to mention being an open model that you can host yourself

US labs must be sweating bullets. Not on tech side, but on finance. They have a pile of debt and VC expectations that count on vast future profitability

Maybe the whole windows/linux thing is an apt comparison? Linux is OS and arguably runs the internet yet windows is still a cash cow. The paid-for models are like windows, the OS models are like linux?

Considering Xi just said they are ALL IN on open AI, yeah, heavy sweating

Don't worry, we're at the stage where we lobby the government for "regulations".

They'll just lobby the government, feds will label them a security threat and ban them from corporate use. The American AI VC will breathe a huge sigh of relief and call it a day.

Nothing says capitalism like writing laws to stop the competition needed to fuel innovation.

I don’t think the timing of Trump’s recent sabre-rattling at China is any coincidence

> Chip Design

> As an early proof of concept, Kimi K3 designed a chip to serve a nano model built on its own architecture. In a single 48-hour autonomous run, K3 built, optimized, and verified the chip using open-source EDA tools on the Nangate 45nm library. Within 4 mm², the chip closes timing at 100 MHz and sustains over 8,700 tokens/s decode throughput in simulation, packing 1.46M standard cells, 0.277 MB of SRAM, and an INT4 MAC array with fused dequantization. A chip built by a model, for a model, reflects K3's long-horizon agentic capabilities.

Absolutely wild.

I had a thought a while back: sell large local models burned onto fused compute / ROM chips. Like cartridges for old game consoles. Slot (or probably plug into USB-C) and go.

It’s an ASIC with the model wired into it so it’s very low power and fast.

I’d buy these. Say $100 for a frontier class model. Maybe more.

Less impressive than you might think here. I can't imagine how small this model must be if it only has 1.46M cells. Maybe 1M? Probably less. Sure the token/s decode is cool, but this size isn't capable of doing anything. For context: this is a 45nm process.

In comparison, Taalas' implementation (which everyone likes to talk about) is in 6nm, is 815mm^2, and only serves a 8B parameter ~4-bit quant model.

So, 815mm^2 in 6nm is roughly equivalent to ~6112mm^2 in 45nm. If we assume everything scales exactly the same, 4mm^2 would be ~1500x smaller. 1500x smaller means at best we're talking about a ~5M parameter model. I don't know how you'd get a 5M param model (with multiple bits) in 1.46M transistors.

groq did an ASIC for llama and now for nvidia. Their cloud service is fast.

> NVIDIA Groq 3 LPU Inference Accelerator > The NVIDIA Groq 3 LPU is the next generation of Groq’s innovative language processing unit. Each LPX rack features 256 interconnected LPU accelerators that, together with the NVIDIA Vera Rubin platform, supercharge inference. Each LPU accelerator delivers 500 megabytes (MB) of SRAM, 150 terabytes per second (TB/s) of SRAM bandwidth, and 2.5 TB/s scale-up bandwidth.

https://www.nvidia.com/en-us/data-center/lpx/

Really feels like end game type stuff - AI designing its own next versions, designing its own chips, etc..

The advancement is slow, but fast - like a plant growing. We really are the boiling frogs now aren’t we?

And the people with eyes wide open are us, and anyone that frequents this site really. Is this Milliways?

I cannot imagine how much context window it needs for this. I am struggling with local AI in Pi to even get small (500-1000 lines of code) programs written without overrunning a 220k context. Doing all this EDA design and testing must surely require millions of tokens of context window right? Maybe they just do a far better job of architecting the work and sub-tasking it to sub-agents, each with their own context, to keep the primary context window from overfilling.

How nano are we talking about here? A single transformer head and a few dense layers?

Check out coresmith.ai to try this yourself

Working with chinese models is giving me a fullfilment sensation. I think that I have enough quality for the work that I need to do and lots of extra tokens to work with. With Claude and ChatGPT I reach the limits fairly easy, but not with OpenCode Go. So I will use Claude once in a while for difficult tasks to see how much better it still is (but use Chinese on a daily basis)

I have been using Deepseek V4 Pro for personal projects and it has been great. I think the $20/mo GPT plan is still the strongest value, but only because you don’t have to pay API prices for tokens.

I benched DS4 flash and Pro vs opus 4.8 xhigh on 16 work-related tasks a month ago across 4 days.

Opus 4.8 came out as a winner by 1 task only where both DS4 pro and flash looped out of "focus". But flash performed as well or better (as in being more thorough) in 13 out if 16.

The way I see it even DS4 flash is as efficient as top dogs and only starts lagging on very vibecodey (generating lots of stuff) or very difficult bugs. But you're really spending low cents amounts for your tasks.

OpenCode Go is a great deal but I recently dumped my subscription because I found myself rarely reaching for it over my Anthropic sub (I can get 40 hours of work a week out of the $20 sub and almost never hit weekly limits). Subscribed to OpenAI as my secondary and I've been really impressed with that too so far.

I expect if they add Kimi 3 to Go the limits are going to be really low since 2.7 is already one of the most limited models and 3 is much larger.

I finished benchmarking[0] it, but it was not fun, it only supports (max) reasoning and the model is quite slow. Apart from a few requests timing out, it also has some issues with tool calling/response format schemas (Moonshot rejected tools.function.parameters with anyOf schema).

It also, for some reason failed to generate either of the 2 coding demos (hamster svg and solar system css animation).

Intelligence-wise, it's between GPT-5.6 Terra and GPT-5.6 Sol. It's ~30% better than Kimi K2.6, but a lot slower and more expensive.

[0]: https://aibenchy.com/compare/moonshotai-kimi-k3-max/moonshot...

Just saw the logs, coding demos failed due to the 5 minute/task timeout. I have increased it and retesting it now.

EDIT: With 10 minutes timeout, the CSS task completed, but the SVG generation task still timed out. Trying again with 30 minutes timeout...

EDIT2: It completed (now in only ~9 minutes). It's one of the best hamsters[0].

[0]: https://aibenchy.com/compare/moonshotai-kimi-k3-max/moonshot...

Amazing to see an open source model already nearing the benchmarks of Fable and GPT 5.6 Sol!

Also very cool to see LatentMoE being picked up by more models (https://arxiv.org/abs/2601.18089)

It also goes to show that Fable/Sol must be 4-5T in size.

Surely it's only open weights?

Anthropic's "durable advantage" theory of US AI dominance is looking pretty silly. There's zero indication that it will be hard for China to keep pace as models improve and start contributing to their own training. Which pretty much invalidates their policy recommendations.

They can't even blame it on distillation this time, unless they want to claim that their own preferred security measures were ineffective in preventing Chinese access to Mythos.

I remember that more than a year ago, when Anthropic and OpenAI started to hide reasoning steps, some were claiming that Chinese models were done, as they could only distill those US models.

I am very curious for the next batch of Chinese models. I have been using DeepSeek and it is nothing short of excellent.

Likely won't improve much. They trained on every text already.

I'm a bit nervous this one isn't going to be open-weights. Any mention of "open" has been struck from the literature for this model (it was present an hour ago). We don't even know active params?

At this pricing, I'll be surprised if it's open.

They will release the full weights by 7/27 along with support in vLLM.

Source: their release blog on WeChat. https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ

Reuters has been reporting that Chinese government is undergoing similar investigation to the US; blocking the export of domestic frontier models. They boil down to "anonymous sources" but it does seem inevitable as the tech gets stronger and stronger.

It's 2.8T, I'm sure they will open the weights but it will only be able to be run on very high end machines.

This does seem like a cash grab. These token rates are crazy. I'll just use GPT 5.6 thanks.

> In our evaluations, Kimi K3 delivers frontier-level performance. Among the models tested, its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol. For the complete benchmark results, see our tech blog. The full model weights of Kimi K3 will be released in the coming days. More details on the architecture, training, and evaluation will be published together with the Kimi K3 technical report.

> K3 pushes the boundary of end-to-end knowledge work. On the GDPval-AA v2 leaderboard, Kimi K3 scores 1687. The benchmark evaluates AI models on real-world tasks across 44 occupations and 9 major industries; Kimi K3 ranks behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8 Max at 1600.

> On AA-Briefcase, Kimi K3 scores 1527, ranking second among all models — behind only Claude Fable 5 Max and ahead of GPT-5.6 Sol Max (1495). AA-Briefcase is a private agentic knowledge-work benchmark developed by Artificial Analysis to evaluate frontier agentic capability in long-horizon knowledge work.

Really good benchmark score it seems. Maybe another DeepSeek moment right here.

> its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol

Pretty sure ranking “second” to two others means ranking third.

> > K3 pushes the boundary of end-to-end knowledge work. On the GDPval-AA v2 leaderboard, Kimi K3 scores 1687. The benchmark evaluates AI models on real-world tasks across 44 occupations and 9 major industries; Kimi K3 ranks behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8 Max at 1600.

This is the same benchmark where Sonnet 5 outperforms Opus 4.8 max.

Like all model releases, the benchmarks aren't going to tell the whole story. All of the open weight models come with amazing benchmark results now. It's hard to believe anything other than that the benchmarks are leaking into (or intentionally included) into training data.

> Maybe another DeepSeek moment right here.

Surely not... What made DeepSeek disruptive was that the cost was 10X lower.

In this case, the cost is about 2X lower the Sol I think?

At 2X, you're pretty close to the error margins due to token efficiency etc...

I'd say this is "on trend" for open models catching up to frontier labs, but its not a "change in the trend" like DeepSeek was IMO.

In my experience, the Chinese models are much more benchmaxxed than their frontier lab competitors, so I'm taking these results with a fairly large helping of salt.

That’s an interesting way to say you’re third. I’m only second to the ten other runners on my local Strava segments.

> In our evaluations, Kimi K3 delivers frontier-level performance

What page does that come from? I'm having trouble tracking it down.

Where are you seeing this write up?

What hardware are they using to run the benchmarks?

Where is this from?

I switched to exclusively Chinese models, mostly Kimi, many months ago. I'll still ask Claude questions that require ambitious real-time web search / worldly knowledge, but for just about anything else, the Chinese models have been so good that I haven't looked back.

I think that those two statements are somewhat contradictory of one another. But in any case, I'm curious about your decision to still use Claude for some questions. Do you find the Chinese models have less worldly knowledge, and if so in what categories? Genuinely curious, I haven't had a chance to try them out very much.

I did the same, although last week we switched back to GPT-5.6 since OpenAI is giving the farm away pricing/reset wise.

Imagine you're a mid sized company and you can host this model locally. Suddenly there are zero reasons to pay a single red cent to the bloodsucking American AI cartel.

Can you host the model for a lower cost per token than you'd pay Anthropic or OpenAI for a similar level of intelligence? I doubt you're beating their efficiencies of scale.

hardware, electricity cost and other extra time consuming deployment, are they joke to you? ROI needs to positive otherwise open models have still BIG COST.

Whether it is "open" or not seems to be in question. While it was initially called an "open" model, it seems that "open" mentions have been scrubbed from website.

Just in case you were thinking of signing up directly with Moonshot to use the service, they appear to train even on API use:

> We may use Content to provide, maintain, develop, support, and improve the Services, comply with applicable law, enforce our terms and policies, and keep the Services safe and secure. Customer who requires restrictions on the use of Customer Content for training or improving Moonshot AI models may contact Moonshot AI to discuss available enterprise arrangements or separate written agreements. Unless otherwise expressly agreed in writing, Customer Content may be used for the foregoing purposes.

https://platform.kimi.ai/docs/agreement/modeluse#4-content

If they want to take my awful prompts and poison their model with it, it's their loss.

For anyone else that believes their input needs secrecy, you need to check the corporate plan of any provider for data protection clauses. Most use the cheaper plans as bait to get more training data and free feedback.

Interesting. OpenRouter classifies the Moonshot provider as ZDR. I wonder whether they have a ZDR agreement or it's a misclassification on their part.

[flagged]

Strictly dominates both Sonnet 5 and Opus 4.8 in both cost and performance:

https://artificialanalysis.ai/models/comparisons/kimi-k3-vs-...

Excited for the deepseek release this week (or at least they announced they'd release this week). Hopefully they also push even closer to SOTA.

> We also further increased the sparsity of the Mixture of Experts (MoE): with the Stable LatentMoE framework, the model efficiently activates 16 out of 896 experts. Together with improvements in training methodology and data recipes, these structural advances give K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively.

Assuming experts are uniformly distributed (I’m really not that familiar with the deep details there), that’s 2800/896*16 = 50 billion active parameters just for the active/expert part. Wild stuff, and I’m glad there’s at least some companies still publishing (and pushing, for open-weight models) total parameter count.

And: It sounds very believable that this would result in efficiency gains wrt. to compute necessary for “good”-quality inference. Does anyone know whether there currently even are any SOTA or near-SOTA models that are dense still?

The blog post is now online:

https://www.kimi.com/blog/kimi-k3

- The blog post is explicitly saying that the model is open; that language was removed from the previously shared link

- It shows benchmarks

I've been playing around with it for the past few hours, and I think it's an amazing model. I'm not sure I could tell the difference between this and Fable in a blind test. The quota in the $100 Kimi Coding plan seems to roughly align with what I get from the $200 Anthropic plan when I primarily use Fable.

Any updated Pareto frontier graphs? https://paraplouis.github.io/llm-pareto-frontier/ is quite out of date now.

Did anyone see on the blog post[0] that it was able to code up an entire GPU compiler from scratch? It looks like it even outperformed triton on some GPU kernels. That just seems insane to me.

Wonder if they’ll open-source this and show how many tokens it cost.

[0] https://www.kimi.com/blog/kimi-k3

The amount of scientific talent in China is astounding, also it's easier to be a fast follower than a innovator.

Instead of limiting models and debating ethics like Anthropic, the edge lab should focus on lengthening the lead on China.

The 2027 Chinese model could be one that beats the US.

Half kidding feature request for HN: Mark all AI related posts so I can filter them out, when I need a pause.

It does seem to have retained the K2 series's creative writing abilities, at least with the prompts I've tested so far.

Based on my usage of Fable on the max plan, Anthropic is going to have to at least double the usage limits for it to be a viable option in the near future. And without Fable, the value proposition quickly drops to zero.

> Among the models tested, its overall intelligence ranks second only to Claude Fable 5 and GPT-5.6 Sol.

> The full model weights of Kimi K3 will be released in the coming days. More details on the architecture, training, and evaluation will be published together with the Kimi K3 technical report.

https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

Only supporting "max" reasoning is weird, their parameters are quite inflexible atm:

    Important limits:

    reasoning_effort currently supports only max; K3 always has thinking mode enabled.

    max_completion_tokens defaults to 131072 and can be set up to 1048576.

    temperature=1.0, top_p=0.95, n=1, presence_penalty=0, and frequency_penalty=0 are fixed; omit them from requests.

    Return the complete assistant message unchanged in multi-turn conversations and tool calls.

    Vision input does not support public image URLs. Use base64 or ms://<file-id>, and make content an array of objects.

    Web search is being updated and is not recommended for production workflows in the near term.

Very interesting to see how Gemini 3.5 Pro stacks up against this new wave of models. Hope they have something similar to a Gemini 3.1 moment soon. Their speciality has always been math and multi modal intelligence and the new models are recently all very coding focused.

For day-to-day programming work, have you seen a difference in the quality of output between (Opus 4.6 / GPT 5.2 / GPT-5.3 Codex) and the current (GPT-5.6 / Fable) that justifies the price increase?

My intuition says that the output quality difference is marginal compared to the change in price especially when taking into account the effects of prompt/context engineering and harness differences.

Essentially: since opus 4.6, working through a model's quirks with prompt/context engineering and harness development will yield significantly better output than just switching models to the latest.

This is the perfect opportunity for anyone who wants to boycott Anthopic and OpenAI!Anyone who cares about the future of Ai has no excuse to continue relying on companies that want to have a monopoly on intelligence.

Account creation with only a phone number or google account is lame.

It's important we now have a recap to the opus 4.8 release where we were threatened with ID verification as "these models become more powerful" and had to pass "verification" to gain full access to the capabilities without having random "cyber" refusals.

I didn’t realize that GPT-5.6 is basically dominating the cost/intelligence Pareto frontier right now, at least for this set of benchmarks. Otherwise it’s only Fable on the very high end and DeepSeek on the very low end. This Kimi model gets close, though.

https://www.reuters.com/world/china/chinas-xi-outline-ai-dip...

Reuters is reporting that Xi is planning to endorse open source/weights AI in a speech tomorrow. This is probably highly relevant to why Moonshot is committing to making Kimi open weights.

The technical blog post is out now, and it's a better top-level link than what we have currently: https://www.kimi.com/blog/kimi-k3

What about the cyber security capabilities? Given that this model is probably not guardrailed like Fable, would we see a wave of zero days?

A really good startup idea right now... Use kimi k3 to reproduce the kimi k3 asic design and start fabbing it immediately. In 12-18 months, start spinning up your own cloud and start competing with the frontier labs ASAP.

Who needs superintelligence with you have 8,700 tokens/s at near Fable levels of performance???

This is like the Bill Gates, Paul Allen moment, but for hardware.

Anthropic needs to IPO and dump on you all’s retirement plans quick

This was only a month and a half delay after Opus 4.8 and Fable 5 spent 18 days in embargo, resurrected with a strict classifier that handicaps it

We’re at endgame

I really wanted to try it out, but I keep getting blocked by rate limit (opencode go). Can you guys get off it for a second so it can process my request? Thanks.

> While its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol

I'm inclined to believe that, however according to their own benchmarks Kimi K3 actually even beats the other two in many metrics, no?

Thank you Kimi. We no longer need to rely that much on Dario and his supreme lackeys to decide what is safe or not for simple tasks.

Looks like open models being months behind is a thing of the past. Now more like weeks.

I just tried this on the monthly $18 plan, having it do a basic task with its 2.7 model and then audit it using k3.

K3 got into some loop trying to run docker and after maybe the 6th attempt ran out of quota for the 5 hour window which represents 20% of the weekly.

I run 200 max and chatgpt pro, but I had to blink at that.

K3 didn't even write out what it was doing or provide any sense for why it was pursuing the execution path it was.

I'm in disbelief that this is a groundbreaking model, and do not think it represents a threat to Claude Code or Codex at this time.

I've playing around in between with Arc-AGI-3 lately. Based on my very quick test prompt, I do not think it will achieve any meaningful score in Arc AGI 3. Not that it was expected to.

I took advantage of their "Token Cup" for the world cup and won 530,000 credits. I believe at the time they said it had to be used in the desktop app, which I have installed. Nowhere can I find any sort of balance or evidence of the 530k other than the Token Cup page itself that say that is what I was given.

Their web chat has almost no settings of customization. Everything they present just comes off as amateurish to me. I trust them less than most Chinese AI companies, which a very low bar.

This is the first time an AI provider makes a trailer[0] for their main release and game development is the first demo and main hook.

Anthropic might dominate general purpose programming,but I think there's enough of a market for a model laser focused on game scripting or tool development for game studios.

I hope it succeeds in serving that audience.

[0]https://youtu.be/bn0atstgavo

I get a quota of GitHub Copilot for free.

From all the models available to me I'm most happy with Kimi K2.7 (given the cost/performance).

The big danger here is the gradual increase in open-weight subscription costs. I use open weight subscriptions, with lower-cost models for 80% of my tasks and GLM-5.2, Qwen 3.7-Max, Kimi-K2.6/2.7-Code for the 20% that need the most intelligence. That lets me maximize the rate-limit the subscription gives (rate limits per model are literally a price-limit-per-token/model). When new/more expensive open weights come in, providers phase out older/cheaper models. Over time we will either have to pay more, or use our subscriptions less.

It goes without saying, but if the open weights become as expensive as SOTA models, there's no point in using open weights. If nobody pays for open weights' development, the development dies out, and we're stuck with a US-controlled duopoly again. Which may be the biggest threat the world has seen from the US since nukes.

Say what you want about these Chinese models but they sure create competition and urgency in the space.

That's a more than 2x jump in parameter count. I know it's not a measure of quality by itself, but it will be interesting how it "scales". Bust it looks like they're gonna be competing with the big boys now, pricing also approaches Gpt 5.6 Terra

Open source Fable/Sol challenger! Interesting to do a release product-first.

https://platform.kimi.ai/docs/guide/kimi-k3-quickstart

What subscription plan for Kimi 3 would be the most cost effective? Most people only talk about API efficiency, but is there any place that evaluates how much you get with the subscription plans?

Kimi 3's Artificial Analysis benchmark scores between GPT Sol and Opus 4.8.

https://artificialanalysis.ai/models

https://www.kimi.com/blog/kimi-k3

"The full model weights will be released by July 27, 2026."

I tried the $40 plan. Seems ok to get some real work done. The model seems quite capable and being able to read the reasoning trace is bonus. It's not the fastest though.

Is K3 marked as a proprietary model because its weights have not been released yet? Were there indications from Moonshot that K3 would or would not be open weights?

Why do most LLMs insist on a login, even for a free trial?

I entered a question to try it, but as soon as I hit enter it wants my phone number for a login. No thanks.

Frontier Intelligence K3 tokenisation of language processing is closer to a market research application.

[1]: NLP: natural language processing

[2]: MLP: machine language processing

The former are subject to iterations of 26 characters, 0-9 integers, recombination of "tokens."

The latter are not object oriented, iterating to the boolean 0-1 disjunction, separating window management programming interfaces from dropping into POSIX.

The pricing on this (it’s… fine?) might say more about the costs of running any frontier/borderline frontier model than any other insight. Either Kimi knows that the cost to run the premier models are being highly subsidized and will right course soon, or the cost for any borderline frontier model at this point is just flattening.

Is anyone using non frontier models as secondary models for coding tasks? What's your setup? I'm on the Max 20x plan for Claude but I still want a secondary, maybe fast model for offloading some tasks and parallel development.

Any recommendation for a cost effective subscription service?

Does anyone have any heuristics on how scaling parameter count actually scales cost to serve? Also im assuming we dont really know the sparsity here?

Is them pricing at Sonnet level actually give us any information at all at how big Sonnet is or is there too much opacity around inference margins?

>Too many people are chatting with Kimi right now. Subscribe to enter a dedicated priority queue!

Seems to only use ≈60% as many reasoning tokens as 2.6. So the price hike is not as bad as it looks.

I'll switch to this for now.

I'm expecting Anthropics reply soon though.

It would be trivial for them to distill Mythos.

I really don't see how the SaaS models will be profitable if this continues, all the money will be in providing consumers with hardware that can run these locally - with the ability to modify the weights and do your own ablation.

This is far too expensive. Why would I use this over a frontier model at these prices.

Not worth it. I have just tried a single prompt in the web interface and it is still not finish reasoning. It thinks too much and often repeats the same stuff over and over.

Combine with the price it will surely more costly than gpt 5.6.

Semi-off-topic, but...

Is the release of this why Google's share price is down 4.5%?

No blog post? Benchmarks?

AA results are out: https://artificialanalysis.ai/models/kimi-k3

GPT 5.6* throw fits on anything even remotely related to reverse engineering, and I'm not ever paying anything more than $20 to Anthropic anymore.

How's Kimi in this area?

I hope this means they stop downgrading my fable requests

I am trying to benchmark it, but it only supports (max) reasoning, and even for simple questions, it takes forever to answer/times out :(

Quite impressed by the result to my first prompt...

How feasible is it to hook Kimi up to do GitHub code reviews? the Copilot quotas got really stingy recently

For comparison, a human brain has roughly 100 trillion synapses, which are sometimes loosely compared to about 100 trillion AI parameters.

This will put Kimi's valuation at least $0.5T

Traditional narrative is that you need tons of traces of actual execution to post-train and get models right. Nobody seems to use Kimi API from Moonshot, I bet everybody is using them on neoclouds/inference providers like Together, Nebius, Fireworks etc. where unlikely they will get traces (in fact, thats the whole promise of these inf providers). How are Kimi models improving so quickly? Is this just distillation (though Sol/Fable just came out so I find it hard to believe)

Some of the gameplay clips coming out from Kimi K3 look insane, going to be a very interesting few months at the end of this year

They're saying kimi3 beat Fable in the AttnRes Kernel Optimization benchmark. What does this benchmark actually mean?

Another deepseek moment? it seems they have fully caught with fable tier of models, and this was a lot sooner than was expected.

I'm not finding this on huggingface yet is and open model or is kimi now a closed model ?

Is there a way to try it without using your Google account or giving them your phone number?

This is why Anthropic is scared of open weight models. Because they are just too good.

This is super exciting. I really need to buy better hardware to try this stuff.

Full benchmarks in Mandarin:

https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ

Translation:

https://mp-weixin-qq-com.translate.goog/s/V4xhEIy8xDXSMDPrPk...

Cheaper then GPT 5.6 Sol (according to their results) ...

Does anyone know how to connect this (web version) to Microsoft Learn MCP?

waiting for - "Running Kimi K3 on X years old hardware".

Holy fuck. This is an open weights model? It's literally trailing the frontier models! This is insane!

It's my dream to own hardware that can run this!

This is too expensive to be a viable model. If it were $5/1m output, it might be another story. At these prices, there's no reason to use this over GPT 5.6.

Can't wait to see the full potential

Wants a phone number...no thank you.

https://macos27.kimi.page

who made this? Looks pretty complete.

I'm curious if they're keeping up mostly due to distillation or how that works. Does anyone outside China know?

Can i trust these benchmarks?

Curious why the thinking mention chatgpt for a moment https://ibb.co/JFdhMN95

I really need to finish my automated model evaluation harness, I can't keep up with this pace

The question remains is it open or not, if it's open I will use it if it's not well I was happily being fucked over by an American tech giant...

China is poking the bubble and it might aswell pop it

I'm disappointed. After all the buzz and benchmarks, I've tested with my personal benchmark that simulates real-world day-to-day specs for agentic coding, following instructions across long time walls, changing several files and code requirements with separation of concerns to build a complete Saas e2e - it reaches a similar rating as DeepSeek V4 Flash.

how much would it cost to host it on AWS for example?

Impressive performance (also the part about Delta Attention, which seems interesting).

Not trying to start a flamewar thread, but isn’t every Chinese LLM censored on certain major political topics? I understand that fine-tuning at least DeepSeek can remove this, but just saying.

really loving this model for brainstorming

Hopefully, gemma5 will have this intelligence next year

do they not have an API? only sub?

whohoo, fable will (have to) stay

So what now? USA blocks Chinese AI. US companies use expansive closed source AI while the rest of the world use cheap self hostable open source AI from China?

But, it's not open models.

Crap, the first open weight model that really feels out of reach when it comes to running it locally at home. :-(

Another Chinese model, another payment wall. Great technology, same Alipay/WeChat problem.

I mean, it's hard not to be impressed by the Moonshot team. Absolutely great work.

With this news - another week of Fable extension for subscribers coming right up.

Now, will they actually release the weights? Seems like Chinese model providers are slowly closing up, like Alibaba's Qwen 3.6 which did release weights (but not the biggest parameter count ones) and none for 3.7.

Silicon Valley and Wall Street are cancer that should be cleansed with fire. If western societies couldn't lit the fire themselves, it is up to the Chinese dragon to burn it.

Aaaaaaaaaand the AI Bubble popped

@dang, since the English blog post is now live:

https://www.kimi.com/blog/kimi-k3

Maybe we should update the link to it instead?

“Open” frontier intelligence? In what sense? Have they announced their intention to release the weights similar to their previous releases?

Kimi doesn't do well on my "ask a trivia question that other AIs get wrong" test.

The question it came up with, "which U.S. state is closest to Africa?" is a pretty standard trivia question without any reason to believe other AIs would get confused. https://pellmell.ai/s/dccdeca69f929f79bc89317035610049

Even GPT-OSS-120b gets this right: https://pellmell.ai/s/1a43dfc7a3baa214aa0fa1b95d2c536a

Reply on news.ycombinator.com