Podcast Notes | SemiAnalysis Breakdown of Kimi k3: China Finally Has a Frontier Model, AI Labs Selling Tokens Could Be More Profitable Than SaaS
TechFlow Selected TechFlow Selected
Podcast Notes | SemiAnalysis Breakdown of Kimi k3: China Finally Has a Frontier Model, AI Labs Selling Tokens Could Be More Profitable Than SaaS
Google should feel embarrassed, while the US government may have directly caused the narrowing of the China-US model gap.
Navigating Web3 tides with focused insightsOrganized & Compiled: TechFlow
Guests: Jordan and Max, SemiAnalysis Analysts
Host: Jordan (SemiAnalysis Internal Dialogue)
Podcast Source: SemiAnalysis
Original Title: [Emergency Episode] Moonshot's Kimi K3 has Arrived! China has a Frontier Model
Broadcast Date: July 18, 2026
Key Takeaways
Moonshot released Kimi K3, surpassing Google and Meta on multiple comprehensive benchmarks, becoming the world's third-best model, second only to Anthropic's Fable and OpenAI's Soul 5.6. SemiAnalysis analysts Jordan and Max dissected the implications of this release in an emergency episode: a 2.8T parameter model served at the same pricing as Sonnet (3/15 per million tokens); if it is comparable in scale to frontier closed-source models, then Anthropic's 10/50 profit margin could be at a mindboggling level.
A sharper judgment comes from Jordan: the frontier gap is narrowing, which he attributes to US government restrictions preventing Anthropic/OpenAI from releasing their strongest models, artificially giving chasers a time window. Open source hasn't truly caught up. Meanwhile, Western open source is completely vacant; no American company's open source model can catch up to China's fifth-best. Model competition is turning into a competition of harness (tool usage) and geopolitics.
Highlights
Regarding Kimi K3's Positioning
- "If you look at the comprehensive rankings of all major benchmarks, today there is a very clear top three: Fable, Soul 5.6, and Kimi K3. They are consistently higher than everyone else, including DeepSeek, and also Google, Meta, and xAI."
- "Google should feel particularly embarrassed. Just in November, December 2025, everyone still thought the AI big three were Google, Anthropic, and OpenAI. Talking to some 'old-timers' today, they still think so, but obviously that's no longer the case."
- "It might be the world's second-best model, because every time I use Fable for serious work I get rejected, pushed back to Opus. Although I'm not sure if it's better than Opus, at least not being rejected is less annoying."
Regarding Frontier Model Profit Margins
- "If Kimi is likely not losing money serving K3 at the 3/15 price point, then Fable being similar in scale but charging 10/50 should completely dispel concerns that AI labs are unprofitable businesses. Selling tokens by API price might be more profitable than SaaS, as of today."
- "From K2.7 to K3, the price increased more than 3 times, from 0.95/4 to 3/15. But I don't think they have much room left to raise prices, because for many tasks GLM or MiniMax M3 is sufficient."
Regarding the US Government and the Gap
- "I believe the reason this gap narrowed is squarely due to the US government restricting Anthropic, resulting in us not getting these companies' truly strongest models. They caught up artificially."
- "We can only access it when frontier intelligence is allowed. This is actually an opportunity for the players ranked fourth, fifth, sixth, and seventh."
Regarding the Western Open Source Vacuum
- "The entire market is still so inefficient, it shocks me that we don't have a single American company that can at least catch up to China's fifth-best."
- "Even if the government doesn't ban Chinese open source models, ordinary large American enterprises are unwilling to feed proprietary data to Chinese open source models. Even if you load weights in an air-gapped data center and the CCP can't see your data, executives won't buy it."
Regarding Harness
- "Testing Kimi K3 made me seriously examine Open Code, Hermes, and Pi for the first time. The harness is completely still part of the product."
- "Some simple things make me choose this model over that: Can it be installed on a remote SSH server? Are the shortcuts easy to use? These details in the harness actually affect where I send tokens, which is where I send the budget."
Regarding "It's Still Too Early"
- "I went to ICML last week, and the week before to the AI Engineer Conference. This is a nominal AI conference, more than 80% of people have never heard of SemiAnalysis. You claim to work in the AI industry, but haven't even read SemiAnalysis? We are still too early."
Is Kimi K3 the World's Third-Best Model?
Jordan: Quick hot take, is Kimi K3 now the world's third-best model?
Max: The answer is a clear "yes". Everyone likes to complain about benchmarks, benchmarks do have problems, but if you take out the comprehensive rankings of all major benchmarks, their direction has always been correct. Today there is a very clear top three: Fable, Soul 5.6, and Kimi K3, consistently higher than everyone else, including DeepSeek and other open source players, and also significantly higher than Google, Meta, and xAI. This is an extremely remarkable achievement for the Moonshot team.
Google should feel particularly embarrassed. Just in November, December 2025, everyone still thought the AI big three were Google, Anthropic, and OpenAI. Talking to some "old-timers" today, they still think so, but obviously that's no longer the case.
Overall I still feel it's not as good as Fable and Soul 5.6. It's a bit funny that they explicitly stated this in their own model release blog. Maybe it's old-school Chinese humility, maybe they don't want to attract US government scrutiny, after all the release of Fable 5.6 was also delayed at the time. But regardless, very impressive.
Jordan: They wrote in the limitations section of the blog: "Although K3 is overall a highly competitive model, there is still a clear gap in user experience compared to Fable 5 and GPT 5.6." My personal usage experience is, it is indeed good, but really slow, which is annoying. It gave me the motivation to try open source harness for the first time. Honestly, I feel I learned more about harness than about the models, because all these models are good enough to complete the basic work I do currently, I hardly find complex tasks it can't do.
This might be the world's second-best model for me, because every time I use Fable for serious work I get rejected, pushed back to Opus. Although I'm not sure if it's better than Opus, the fact of not being rejected is itself less annoying. But when paying by usage with API key I don't get rejected, only when using web console or deep research do I hit limits. They obviously don't have enough GPUs to serve the demand this model brings. Previously this problem was solved through open source strategy, throwing out weights for others to serve. But this time they haven't thrown weights yet, they said 10 days later.
Why Delay Open Sourcing Weights by 10 Days
Jordan: What do you think is the strategy behind this delay?
Max: To be clear, this is all my pure speculation. A big reason might be they need to give inference engine teams like vLLM and SGLang enough time to ensure they can serve this model with high performance. If they throw out the weights today, everyone is serving but only gives you 20 tokens/second, that's terrible for their brand capture. They now have a great opportunity to get a lot of PR and adoption, if performance drags them down upon release, it would weaken the momentum.
Another possibility is they are negotiating licensing partnerships with inference service providers like Together AI, Fireworks, Nebius, Groq, letting them use latest chips like GB300 to serve incremental capacity. These two points are probably the main reasons for the 10-day delay.
Exorbitant Profits of Frontier Models
Jordan: Let's talk about model architecture. It's 2.8T parameters, won't fit in B200, you need B300, GB300 or AMD MI355X to serve on a single 8-card HGX server. Of course you can do cross-node pipeline parallelism, but that will seriously affect performance. So only those with latest chips can serve this model.
Returning to the topic about Google you just mentioned, this model reached frontier competitiveness at 2.8T parameters, this actually gives us some clues about how large closed-source frontier models are. If they are comparing with 10T parameter models, that's even more embarrassing. We have to assume it's in the same magnitude as Soul, Fable.
Max: Yes, you're right. I still believe in the capabilities and acumen of the Anthropic research team. If someone on Twitter says current closed-source models have 10T parameters or something, if that's true, dude should pack your bags, Nvidia stock should drop 50% tomorrow, it's all over.
I'm fairly certain Kimi K3 isn't much smaller than current leading closed-source models, even possibly slightly larger. If this is true, this further confirms a point we've been emphasizing at SemiAnalysis: The profit margins of these closed-source labs are absolutely at a mindboggling level. If Kimi is likely not losing money serving K3 at the 3/15 price point, same pricing as Sonnet; while Fable is similar in scale but charges 10/50, then concerns that AI labs are unprofitable should be completely dispelled. Selling tokens by API price might be more profitable than SaaS, as of today.
Jordan: No employee costs, only GPUs. What about compared to previous pricing? You said 3/15, but the previous version Moonshot direct pricing was 0.95/4, so from K2.7 to K3 the price increased more than 3 times. How much room do they have left to raise prices?
Max: I don't think they have much room to push up. Even at the 3/15 price point, many people will feel it's too expensive. Their tasks are sufficient with GLM or MiniMax M3. There is an interesting divergence here: like us at SemiAnalysis who don't care about burning Dylan's tokens, will continue to use Fable for almost everything; while extremely cost-sensitive ones, like Tesla, Uber who only use $200 tokens a week, will go for GLM pricing tier. Then who are the users that will actually switch to Kimi K3? Probably a group of people who philosophically love open source and want to support new models. Will large enterprises actually adopt this model, I wouldn't be surprised if the answer is no.
New Architecture and Next Steps
Jordan: This is a completely new architecture. 2.8T parameters, has Kimi's delta attention, potential residuals, stable latent, basically an amplified version of previous models, about twice as large. Previously K2.5 we saw Cursor use it for composer, based on continued pre-training and MRL, then came 2.5, 2.6, 2.6.7 etc. checkpoints. This is a new base model, but already quite complete to use,没有出现 rough edges common in raw models. What's next? When will 3.1 come out? Will pricing change? Will there be a composer based on K3?
Max: There probably won't be a composer based on K3, the Cursor folks have already decided to train their own models from scratch. As for K3.1, K3.2 etc.,估计 next one or two months will see two to three updates, just continuing post-training. Pricing I guess will remain unchanged, because they can't run on new hardware in the next two or three months, no throughput improvement to lower prices. Maybe some awesome kernel engineers can press cost down to DeepSeek V4 levels, but I'm skeptical about 3T parameter models achieving this. Current GLM and MiniMax pricing might already be the limit serveable by 1T to 1.5T models.
Will Open Source Catch Up to Closed Source?
Max: The more interesting question is whether the gap between open source and closed source will continue to narrow, can open source truly reach parity at frontier level. If this happens, the impact on our entire industry is huge. What do you think?
Jordan: My view is that the gap has narrowed now, the reason is squarely due to the US government restricting Anthropic, resulting in us not getting these companies' truly strongest models. They were caught up artificially.
We can see the comparison between Mythos and Fable. I can't use Mythos, and I have to beg to occasionally use Fable. 5.6 Soul, we internally judge it's not the largest model OpenAI has trained, not as large as 4.5. They have a larger one in hand. The result is, we can only access it when frontier intelligence is allowed by the government.
This is actually an opportunity for the players ranked fourth, fifth, sixth, and seventh, to unleash everything within a certain upper limit, start grabbing user share, but never touch the true frontier. I think the frontier might take another big step forward by late this summer, or political winds might change a bit. Also possible we start finding modalities beyond coding, letting them truly explore those fields.
Mentioning Thinking Machines' Inkling release in passing, native audio input I think is very interesting, a signal of the future.
Max: Regarding Inkling, the West is indeed very, very, very lacking a non-terrible open source model. The market is still so inefficient, it shocks me that we don't have a single American company that can at least catch up to China's fifth-best. On one hand, the US government completely banning Chinese open source models might just be a matter of time. On the other hand, even if not banned, ordinary large American enterprises are unwilling to feed proprietary data to Chinese open source models. Even if you load weights in an air-gapped data center and the CCP can't see your data, executives won't buy it. Many enterprises care about token budgets, and are only willing to run Western models or non-Chinese models. Inkling is the best Western OSS we can get now, but still far from open source frontier, which shocks me.
Jordan: Previously it was Neotron, now it's Inkling. I think Inkling's strategy has two opportunities: One, they must be better than most Chinese open source, only then can they enter the game. Two, they also have to be better than frontier labs' secondary, tertiary models, better than Sonnet, because you can get near-frontier intelligence using Bedrock or Foundry, using closed-source secondary models to save money. I've never quite understood the Western open source "helping people save money" angle. Pushing models into Fireworks, Together, Base10 etc. ecosystems is indeed good, but the bulk of the market is at the government level.
Built for Chinese Domestic Accelerators
Jordan: Another worth mentioning, the K3 blog mentioned quantization was done during the SFT stage, natively using MXFP4 and MXFP8 for weights and activations, official statement is "broad hardware compatibility". Do you think Moonshot cares about any other hardware?
Max: I have a list of 11 types of Chinese accelerators, you should subscribe to SemiAnalysis Accelerator Model to learn more. Huawei Ascend, Baidu Kunlun, Cambricon, Moore Threads, various chips appear in papers, and also seen in code. Running frontier models on domestic accelerators is already China's national priority. If we were still saying Google is a frontier lab at the end of 2025, then now we also have to call Moonshot a frontier lab.
Jordan: Speaking of digressions, my dad is on a business trip in China, he said the hotel he's staying at is fully booked, because Xi Jinping is about to arrive in that area to give a speech about AI being China's top priority. Much of what you said is correct.
Harness Is the Product Itself
Jordan: The biggest realization I have while using these models is, first, it's getting harder to distinguish between using absolute frontier models plus max thinking mode, and using medium effort. In daily tasks I really can't find things these models can't handle. My behavior by default is to turn on the largest, hardest mode, because I don't care about Dylan's budget.
But there is a level where the harness itself is part of the product. Testing Kimi K3 made me seriously examine Open Code, Hermes, and Pi for the first time. The harness is completely still part of the product. Some simple details make me choose this model over that: Can it be installed on a remote SSH server? Are the shortcuts easy to use? Can previous commands be edited? These small details in the harness actually affect where I send tokens, which is where I send the budget.
Max: Many people talk about token budgets, but from your workflow description, even for tasks GLM can handle I'm willing to route to Fable using max intelligence, because ROI is still worth the price. Benchmarks say many tasks can migrate to GLM, but you are still willing to stay on Anthropic or OpenAI models.
Jordan: Basically. But I use many Slack bots, I don't know what model runs behind them. For example Perplexity's Slack integration, if it starts routing to K3, routing to GLM, routing to Sonnet, I actually don't care. Previously I only realized how much OpenAI models accounted for when looking at usage, because this was its own decision. The justify for that part is the harness deciding.
Max: This is actually an entry point for outcome-based pricing. If a lab does outcome-based pricing, they might get more than 95% gross margin, because tasks you are willing to pay stable pricing for today, actually can be done for a fraction.
Jordan: Second point, I don't think these labs have run out of ideas. They can continue training amazing models, to hit coding side RSI, but not release to us, maintaining their "permanent underclass". They can continue distilling, giving us a taste, while continuing to explore other uses, like video generation, audio-to-audio, deep research, these are not quite like coding. Robotics and world models are a simple direction; what if Anthropic shifts its goal from knowledge work to physical labor? I don't believe they can't build a sustainable good ROI business using the world's greatest technology.
We Are Still Too Early
Max: Even without talking about these, I use these models very heavily every day, my friends who are software engineers use ten times less than me and spend ten times less. A person using Fable and a person using Sonnet use the same amount, but the person using Fable spends ten times the money, 90% gross margin,撑起 bulk of the business. Once those people start using larger models, use more, demand will only be larger, models don't even need to get better. Then I still have to talk to neighbors who don't do tech, among them I'm definitely 0.1% or even 0.01%, maybe there's still 1000 times growth space. Returning to Masa-san (Masayoshi Son)'s "Golden Goose Exponential Curve".
Jordan: What she said "it's still too early" is completely correct, this is also why I feel Kimi K3 won't slow down Anthropic and OpenAI's net new ARR. Even if today some people using Fable and 5.6 不可劝地 switch to Kimi K3, this group of people will be completely overwhelmed by those who haven't seriously tried this technology yet. Those people are discovering new high ROI use cases every day, they will still default to using 5.6 Soul or Fable 5 to unlock these new scenarios. You won't see ARR growth slow down.
Max: Think about how many people haven't subscribed to this podcast, haven't followed SemiAnalysis. I went to ICML last week, the week before to the AI Engineer Conference, this is a nominal AI conference, more than 80% of people have never heard of SemiAnalysis. You claim to work in the AI industry, but haven't even read SemiAnalysis? We are still too early.
Jordan: This counts as an ego check for yourself, Max, calm down.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
深潮TechFlow
