Is Kimi K3 Really Fable Class?

0:00Today on the AI Daily Brief, did we actually just get a fable-level open model? The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:15All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Blitzy, and Airtable. To get an ad-free version of the show, go to patreon.com slash aidailybrief, or you can subscribe on Apple Podcasts. And of course, to learn more about sponsoring the show, send us a note at sponsors at aidailybrief.ai. All right, friends. Well, today we are talking about Kimi K3. The month of models continues, and today we're going to try to figure out just how significant this one is. At first glance, there are some very significant and bold claims being thrown around, but we're going to unpack what's real, what's not, and what the implications are.

0:54And to understand this, or frankly, any frontier Chinese model, you have to put it in the context of the way that the US market sees the AI race. Since we're coming to the end of the World Cup, let me plumb for a soccer analogy. When it comes to our lead on China in terms of advanced models, we very much have one to zero type of energy. What I mean by that is that there's no doubt that we're in the lead, but the scoreline is not something that anyone is particularly comfortable with. The US tends to act like it can feel China coming up on our heels, pressing their advantages and trying to find the equalizer.

1:26In other words, despite being in the lead, it can sometimes feel like we're the ones hanging on. And by the way, for any of you Three Lions fans out there, I am so sorry to use this analogy in this particularly difficult moment. In any case, you can see examples of this feeling of China nipping at our heels spread throughout the last couple of years. The best example, of course, was when DeepSeek R1 was released, and it ripped hundreds of billions of dollars of market cap off some of the leading companies, including NVIDIA, which had the biggest one-day fall in dollar terms in stock history. And yet that DeepSeek moment set the tone for all the future quote-unquote DeepSeek moments that would come in more ways than one.

2:00What I mean by that is that not only was it a moment where the market freaked out about China having caught up or even exceeded US capabilities, reacting quite severely in market terms, but it was also just pretty meaningfully overblown. It's not that DeepSeek's R1 wasn't impressive, but a big part of the reason that it seemed so impressive was that it was democratizing access to a technology that had thus far been locked behind a paywall when it came to companies like OpenAI. The model itself was actually still pretty meaningfully behind what the leading Western labs were doing, but that didn't change its ability to create some pretty significant psychological scars.

2:33Now, ever since then, we have been having many deep-seek moments at a fairly regular clip. The most recent one came when Fable 5 was locked down as per government order, when Z.ai's GLM 5.2 came out, leading to not only positive reviews on Twitter, but also this piece from the Wall Street Journal, which was printed and slapped on desks all over Washington, D.C. The article was called China Has Matched Anthropic in Cybersecurity Resetting AI Race, and as we discussed a lot then, was once again another example of the narrative being fairly overblown, but continuing to be persistent as something the US was worried about.

3:07For the last month, really ever since Fable 5 was released, there have been debates around how long it will take for Chinese companies to have a Fable 5 class model. In the middle of June, Elon Musk predicted Q1, to which the founder of Z.ai responded, won't take that long. And so that was the setup coming into the announcement of Kimi K3. Now, Moonshot's Kimi models have been some of the most popular when it comes to Western users using models from Chinese labs. In fact, as we've been discussing fine tunes of open models like Cursor's Composer 2.5, they tend to be built around a Kimi base. On Wednesday, the Kimi K3 teaser started coming in a serious way.

3:44AI leaker Leo Synthwave wrote, I think Kimi K3 is going to shock some of the Chinese are eight months behind the Western frontier people. And then on Thursday, we actually got the model. Let's talk first about the specs. K3 is a 2.8 trillion parameter model, placing it in a class of its own when it comes to open models. Until now, only a small handful of open models were even in the trillion parameter class, beginning with the first version of Kimi K2 last summer. DeepSeek V4 Pro released this April is a 1.6T model. Xiaomi's Mimo V2.5 Pro is a 1T model. And Thinking Machine's Inkling model released this week is just shy of 1 trillion.

4:20And that's it. GLM 5.2 from Z.ai, the model that got all that bluster that we were just talking about, was only a 744b model. Now proprietary models don't publish their parameter counts, but K3 is likely to be around the same size or maybe a little bit larger than Opus 4.8, but certainly not as big as Fable. In other words, this is a scale of model pre-training that we haven't seen demonstrated by the Chinese labs before. As for features, KimiK3 supports a million token context window and native image inputs alongside text. It uses a mixture of experts architecture, which has become standard for both open source and proprietary models since it was introduced by DeepSeq.

4:56And the benchmarks, well, the benchmarks look incredibly strong. Close to a match for, and in some cases exceeding, Fable 5 and GBT 5.6 sold. On coding benchmark DeepSwee, K3 scored 67.5, which put it 8.5 points ahead of Opus 4.8 and a half point ahead of GBT 5.5. It's 2.5 points behind Fable 5 and 5.5 points behind GPT-56 Sol. For Terminal Bench 2.1, K3 scored 88.3, placing it just a half point behind 5.6 Sol and a few points ahead of the leading models, including Fable 5. In general, at least according to the benchmarks, the model looks pretty close to state-of-the-art encoding and clearly ahead of Opus across the board.

5:39That story is pretty similar for agentic work. K3 scored 1668 on GDPVAL AA, around 70 points ahead of Opus 4.8 and around 90 points behind Fable 5 and 5.6 Soul. K3 is state-of-the-art in Browse Comp and Automation Bench, beating its Western rivals. And on AA Briefcase, which focuses more on long horizon work, it was very close to Fable 5 state-of-the-art performance and slightly ahead of GPT-5.6 Soul. Artificial analysis confirmed the benchmarks highlighted by Moonshot, giving K3 an overall Intelligence Index score of 57. That put the model in third place, three points behind Fable 5 and two points behind 5.6 Soul.

6:16It landed one point ahead of Opus 4.8 and two points ahead of 5.6 Terra and GPT 5.5, which have the same score. K3 is clearly the strongest open model ever on AA's benchmarks, six points ahead of GLM 5.2, which is a huge gap in the context of the Intelligence Index. Another point emphasized by AA was how big of a jump this was from Kimi 2.6. Moonshot picked up 13 points with their new release and moved from 16th place to 3rd. In other words, this is clearly a very strong new pre-training run that could give a solid base model for future iterations as well. Now, AA did also highlight that cost per task had tripled compared to K2.6, which is something that we'll come back to in a little bit in terms of the implications.

6:58It should be clear that this still wasn't a particularly expensive model for the benchmark run at$0.94 per task, compared to$1.04 for 5.6 SOL,$1.80 for Opus 4.8, and$2.75 for Fable 5. But it is of course staggeringly expensive compared to, for example, ultra cheap$0.04 per task for DeepSeq v4 Pro. On the Vals AI Index, K3 did even better. Vals tweeted, Kimi K3 is the number two overall model on the Vals Index, surpassing GPT 5.6 SOL. K3 improved 20 percentage points over its predecessor in less than three months. It is also the first open-weight model of its size. Moonshot is a testament to the accelerating capabilities of open-weight models, which are now competitive with the closed-source frontier.

7:39People quickly dived in to give their examples of what K3 could do, starting with the teaser video itself, which Moonshot claimed that K3 had created on its own, including clip selection, cuts, and audio sync. Moonshot credited K3's native multimodal architecture, which can reason across text, audio, and video, as being able to do this work. Moonshot also gave a bunch of proprietary demos, around game development and 3D digital creation, which is what a lot of people's first tests were for the model as well. Cognition ambassador Justin Goria wrote, KimiKate 3 is a new milestone. K2, 2.5, 2.6, and 2.7 all had the same base model.

8:13Maybe K3 is building on a new one. It's incredible at 3D and front-end tasks. And to prove his point, Justin shared a single-file HTML Minecraft clone. Chetislua shared a one-shot generation of a Voxel Statue of Liberty, which was one of about a million similar examples that were flooding onto Twitter over the course of Wednesday night into Thursday morning. People really love remaking old games as a test. AnyAPI.ai asked K3 to build a 3D Duck Hunt remake in a single HTML file, which it did in around 130 seconds at a cost of 14 cents. Ethan Mollick gave K3 his shader test, create a visually interesting shader that can run in Twiggle.app and make it like an infinite city of neogothic towers partially drowned in a stormy ocean with large waves.

8:52Very good model, he said. Not SolMax or Fable, but great for open weights. And yet there were plenty who were willing to say that it wasn't just great for open weights, but actually was challenging the state of the art. Alex Finn wrote, I was wrong. I said we were a year away from Fable 5 on our desk. That day is today. An open model better than Fable 5 in some benchmarks just dropped. Better than ChatGPT 5.6 on Frontier Suite. Better than Fable 5 on AutomationBench. This fundamentally changes the AI race forever. If people can start running Fable 5 on their desk, unlimited and for free, they're not going to pay thousands for subscriptions.

9:27Now let's be clear, will the hardware you need to run Kimi K3 be attainable for most? No, it won't. It will require multiple very expensive NVIDIA chips or a bunch of Mac Studios. But look at the slope, not the Y-intercept. Over the last year, local AI has become significantly more efficient and required way less compute. The smartest brains in the world are all attacking the compute problem as hard as they can. This is another step in that direction. Within a year or two, you'll be able to run this on a Mac Mini. Local AI has arrived, and it's not going anywhere. Analyst Max Weinbach asked Kimi K3 to create an agent swarm to recreate macOS 27 with real Liquid Glass and native apps in a web browser, and after hours of running on its own, it did exactly that.

10:06Dragos Ruh wrote, I asked Kimi K3 to find my apps in the App Store and to give me feedback on them. It found everything, and this is the most interesting thing, it found ways to circumvent geogating. App Store shows different pages for China. Eventually, it found a way around this, identified IAPs and simulated traffic. I didn't use it yet to code, but so far it's the most complete and polished experience with an open source model. AI early adopter Daria Anutmaz wrote, I just created this interactive website and its entire content by KimiK3. I'm absolutely blown away. This is a potentially new immune engineering strategy for cancer treatment.

10:38KimiK3 conceived 100 % of the scientific content and the design. Jeffrey Emanuel unleashed it to review his 1.5 megabyte markdown plan for his Frankengraph DB project, saying the plan has already been exhaustively reviewed by both Fable Extra High and Sol Ultra, so the low-hanging fruit is gone now. After about 45 minutes, he said, I think the results are wildly impressive here. He pointed out that no other open model could actually provide substantive and correct feedback on a plan that had already had many tens of millions of Sol and Fable tokens behind it. Now he went on to clarify, to be clear, I don't think K3 is as good as Fable or Sol, but it's very strong and not so far behind those models, and most important, it's different.

11:18Different architecture, different training data, different training procedure, different attention mechanism, etc. Which means it will blend well with Sol and Fable and can help find problems that both of those models missed. And when K3 makes mistakes, Sol and Fable can correct for that and ignore the wrong parts. Now another example of K3 not just nipping at the heels of Fable 5 came from Arena.ai, which has KimiK3 as now their number one in the front-end code arena, a 17-place jump from Kimi K2.6, from number 18 to number one. In front-end overall, they said K3 ranked number one in six of seven domains, brand and marketing, reference-based design, data and analytics, consumer product, simulations, and content creation tools.

11:57The only area that it landed number two was in gaming, and that was behind Fable 5. On NextJS.org slash evals, Vercel CEO Guillermo Roche noted Kimi K3 is the best-performing model ahead of Fable, reaching a comparable success rate in less time. Guillermo continued, this is the first time that an open model is ahead of all proprietary ones for this comprehensive web engineering benchmark. He did caution, benchmarks don't always tell the full story. But he did say this is an important signal, adding to mounting evidence that this could be a breakthrough moment for open models. And for some, the vibes followed.

12:29Signal wrote, one last thing I will say before I go to bed, Kimmy is insane at coding, as good as, if not, maybe even better than Fable. It also seems to have excellent design sense as well. It misspells English a bunch though, but animations are crisp too. I can't believe I need to go to bed. I could probably stay up the entire night with this. I feel like a kid again. Bloomberg's Joe Weisenthal even retweeted the Code Arena results writing, is this why the NASDAQ is dropping?

12:57One of the most important AI questions right now isn't who's using AI, it's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com slash US slash sophisticated.

13:35That's kpmg.com slash US slash sophisticated. One thing I keep seeing in enterprise AI, companies hedging across every cloud, every model, every framework, or paying a GSI for a pilot that never ends. The team's actually shipping, they've picked a lane, and they move fast. That's one of the reasons I like today's sponsor, Robots and Pencils. They've gone all in on AWS. They're an advanced tier and AWS pattern partner, and they ship production AI co-workers in 45 days. That's led to them doing some of the more interesting work I've seen on AI co-workers. And by that, I'm not talking about chatbots.

14:08I'm talking about actual agentic systems that sit inside a business architecture and do real work. That kind of focus matters if you're an enterprise leader trying to get something real into production, or an AWS rep trying to move a customer from interested to deployed. Request an AI briefing at robotsandpencils.com. One conversation with robots and pencils, and you'll know. Blitzy is driving over 5x engineering velocity for large-scale enterprises. A publicly traded insurance provider leveraged Blitzy to build a bespoke payments processing application, an estimated 13-month project, and with Blitzy, the application was completed and live in production in six weeks.

14:42A publicly traded vertical SaaS provider used Blitzy to extract services from a 500 ,000-line monolith without disrupting production 21 times faster than their pre-Blitzy estimates. These aren't experiments. This is how the world's most innovative enterprises or shipping software in 2026. You can hear directly about Blitzy from other Fortune 500 CTOs on the modern CTO or CIO classified podcasts. To learn more about how Blitzy can impact your SDLC, book a meeting with an AI solutions consultant at blitzy.com. That's B-L-I-T-Z-Y.com. This episode of the AI Daily Brief is brought to you by HyperAgent, where you run fleets of agents your team can manage together.

15:18New users get$1 ,000 in inference. Forget local agents and chat workflows waiting on your laptop to be prompted. Hyperagent deploys always-on agents in the cloud, doing real work across the tools your team already uses. Marketing's agent turns competitor moves into landing pages. Sales's agent enriches leads, drafts emails, and updates the CRM. Ops agent chases the paperwork and tracks the budget. Every agent has access to shared context and follows your rules about scope and approvals. It's time you add agents that feel like teammates. Hire yours at Hyperagent, built by the team at Airtable. Claim your$1 ,000 in inference at hyperagent.com slash AI Daily Brief.

15:57But at this point, of course, it's worth taking a big old breather. As I was reading all of this yesterday, there was no tweet I related to more strenuously than Dan Shipper from Every who wrote, we will vibe check Kimmy K3, but I am extraordinarily skeptical of claims it's as good as fable. AI engineer Divium, meanwhile, put that skepticism in historical context. He wrote, Kimi K3 is getting a lot of hype, and it is the same cycle we see with every new Chinese model. People fall in love with polished UI demos built in cheap HTML files, fake operating systems, car games, Minecraft clones, flashy dashboards.

16:34And to be fair, Kimi K3 is genuinely excellent at UI work, probably better than some top models. But that is not an accident. These models are heavily optimized for the exact visual coding tests people keep recycling online. The real test starts when you put them inside an actual codebase, understanding existing architecture, tracing a real bug, and fixing it without hallucinating half the project. I gave KimiK3 a debugging task. It could not identify the bug, could not fix the issue, and started inventing explanations. I gave the same task to Fable5 and GPT-5.6 and Medium Reasoning. Both found the problem in one shot at the fix.

17:09That is the gap nobody wants to discuss. K3 can build a beautiful shell. Frontier models can understand what is happening underneath it. Do not confuse a gorgeous demo with real engineering ability. Now Divium is of course here talking specifically about coding, but this idea that Chinese models tend to focus on both maximizing the benchmarks as well as satisfying the early adopter tests that people love putting these models through is absolutely and incontrovertibly true. And pretty soon, as people dug in deeper, we did start to see a few less successful results being shared. Red Kendall wrote, KimiK3 failed at the LavaLamp benchmark.

17:445.6 produced a much cleaner, more realistic result, while K3 struggled with the shape, motion, and overall visual quality. K3 still seems weaker at precise visual generation. Ethan Mollick wrote, Kimmy K3 cannot write a good murder mystery, though neither can any other model, that remains the jaggedest of frontiers. They both make things too obvious and too obscure, and cannot foreshadow to save their artificial lives. And on Bindu Ready and Abacus' benchmark LiveBench, Bindu writes that K3 closes the gap but still ranks behind other frontier models. Indeed, on their benchmark LiveBench, they found that while K3 was the best open source model, it was below not only Solon Fable but also 4.8.

18:22Also in practice, Bindu writes, Kimi spins a lot and costs as much as Opus 4.8 for near Opus class problems, and it's also much slower. And this is the other thing that people started to quickly point out. Specifically that it was incorrect to lump this in with DeepSeq style models, which are super cheap and easy to run. while this model is an open-weights model, it's a big, lumbering, slow, and costly open-weights model that sits alongside the others not only in terms of performance, but also in terms of its cost and difficulty to run. Ryan Fedosuik writes, K3 is a historic moment in the development of AI, but it's not exactly downloadable to your laptop.

18:58In fact, few organizations will be able to local host this capability. Just to hold a 2.8 trillion parameter model in silicon, you'd need the rough equivalent to 44 Mac Studios or 15 Blackwells. a whole NVL-72 rack to the tune of hundreds of thousands of dollars of compute spend. This is why compute will continue to serve as a soft barrier between the capabilities of individuals and organizations. And when you look at costs, certainly K3 is less expensive than the Frontier Western models, but not by the type of margins that I think most people think when they think of less expensive Chinese models.

19:29Jemin Ball points out,

19:49Cognition's Jeff Wang writes, Chinese open source is no longer six months behind, but it's also no longer 10 % of the cost either. In discussing a run of his SVG Pelican benchmark, Simon Willison wrote, K3 only has one reasoning effort right now max, and it shows. The model consumed 13 ,241 reasoning tokens to output 3 ,417 tokens of response. This is expensive. Indeed, at the moment, there's a lot of discussions of wildly expensive simple tasks and failed long horizon runs, so it might just take a little while to get a true sense of how token efficient this model is once the teething problems are figured out.

20:24As an example, Shreyas Mididoti writes, Very mixed results with Kimi K3. Really, really good in general, but gets to dumb reasoning loops burning tokens wasting money. Henry writes, Versus 5-6 Sol, K3 uses over twice the tokens and costs around 40 % more per task for a slightly lower AA intelligence index score. There's also the question of speed. Mark Erdman wrote, Oof, Kimi K3 is slow, eh? Just ran the same prompt through Fable, Sol, and K3 all on Medium. K3 took at least 2-3 times longer. Tons of time spent on thinking. Then it failed partway through generating the HTML output. Dax from OpenCode wrote, So obviously N equals 1, but anecdotes matter a lot here.

21:03I gave KimiK3 and Sol the same task, a simple issue with hovers in the TUI being the wrong color. Sol found and fixed the issue with 30 cents of spend. Kimi got up to a buck and started reading my database before I interrupted it. Indeed, even some of the people who had had good initial results also found some places where KimiK3 failed. Ethan Mollick wrote, A note of caution. I will say that when doing some complex statistical auditing of some of my prior academic work, K3 Max messed up in a bunch of ways, including misapplying statistics and applying some stuff badly. Sanyam Satya wrote, We ran a front-end eval from an in-progress internal benchmark.

21:36K3 is not at the same level as current frontier models like 5.6 Sol or Fable 5. It's closest to Opus 4.7 on this eval, so three months behind the frontier. An impressive result nonetheless, but evaluating frontier models is going to increasingly require very high-taste and in-depth domain expertise. Now, it's important to note that even with all those critiques, there was no one saying that this is a bad model. It's just a natural second reaction after hearing that it was in the same class as Fable to go figure out and test whether that was actually true, which many found just was overselling it at least a little bit.

22:08Now, one interesting thing that didn't come up very much is dismissing K3 as just some distillation. Pim DeWitt did point out that in their tests, when they said, hi, Kimmy, can you tell me about the current weather conditions in NYC. The model responded, just a quick note, I'm actually Claude, not Kimmy. But otherwise, most people thought this was a moment to get beyond the distillation arguments at least a little bit. Mixpanel founder Suhail writes, every single credible researcher I've talked to these past few weeks has said the distillation from Chinese labs is way over exaggerated. Narrative violation, maybe the Chinese are actually good.

22:40Tyler John wrote, I wish we didn't pretend Chinese AI development is a binary matter of this is all distillation versus Chinese companies innovate. Chinese companies innovate. Also, their model capabilities, including those of K3, are strongly bootstrapped via distillation. Both points matter. Nathan Lambert writes, at this point, the distillation arguments need to die and understand that China is also very good at building models. Carnegie Mellon PhD, Xinyu Yang, who now works at Moonshot, wrote, why can Kimmy ship K3? Let me tell my story. Earlier this year, I left academia for industry. I talked to a lot of companies along the way.