Is This the Biggest AI Release of 2026? (China’s New De… — Transcript

China may have just created its biggest AI shock since DeepSeek, but this time the story is much larger than one powerful new AI model. On the same day that Moonshot AI unveiled a giant new model called Kimmy K3, Xiinene Ping stood on a stage in Shanghai and told the world that China doesn't plan to follow America's rules for artificial intelligence. It wants to help write the rules, spread its own models across the developing world, [music] and build a new AI order with China at the center. And Kimmy K3 gives that speech real weight because this isn't another model that looks impressive only inside a company presentation. Now, K3 has 2.8 trillion parameters, making it the largest open-weight AI model ever announced. Moonshot says it's the first open model to approach the 3 trillion mark and its full weights are expected to be released on July 27th. Companies and researchers will then be able to download it, modify it, and run it on their own infrastructure instead of being permanently locked inside Moonshot's app. But the size isn't why this release is sending shock waves through the industry. And the real reason is that Kimmy K3 appears to be dangerously close to the best American closed models. And in some tasks, it's already beating them. Moonshot says K3 can compete [music] with Anthropic's Fable 5 and outperform Claude Opus 4.8, GPT 5.5, and parts of GPT 5.6 in demanding coding work. And independent testing actually tells a similar story. Artificial analysis gave K3 a score of 57 on its intelligence index. Claude Opus 4.8 scored around 56, GPT 5.6, Six, Terra scored 55, and Gemini 3.1 Pro landed at about the same level as K3. So only Claude Fable 5 and GPT 5.6 Soul remained ahead and even there the gap [music] was only two or three points. For years, the standard belief was that open models were 6 to 12 months behind the best closed systems. More recently, people started saying Chinese labs were perhaps 3 to 6 months behind. One Reddit commenter described the new reality more bluntly. Maybe they're six days behind. Now, the most dramatic result came from arena.i's front-end code arena, where developers ask models to build real websites and interfaces, then vote on which result is better. Kimmy K3 entered at number one with 1,679 points. Claude Fable 5 scored 1,631. GPT 5.6. Saul scored 1,618. K3 finished first in six of the seven areas tested, [music] including brand and marketing work and data dashboards. Moonshot's previous model, Kimmy K 2.6, had been sitting in 18th place. K3 jumped 17 positions in one generation and went straight to the top. Even Arena's CEO called it potentially the biggest AI release of the year and suggested it could mark the moment China moved ahead of the United States in at least one major part of the model race. Now, Moonshot isn't pretending K3 wins everything. The company admits it still trails Fable 5 and GPT 5.6 Saul in overall user experience [music] and some broader tasks. But K3 doesn't need to dominate every category to be disruptive. It only needs to make companies ask an uncomfortable question. Why keep [music] paying premium prices to American providers if an open alternative is almost as capable? So, Moonshot built K3 mainly for long-running work, especially software development. And this isn't a model designed only to answer one prompt and stop. It's meant to inspect a large codebase, create a plan, use tools, [music] make changes, check whether they worked, and continue for hours with limited human help. It can also look at what appears on the screen. Moonshot calls this vision in the loop. The model writes code, looks at the result, notices what's wrong, changes the code, and checks again. That makes it useful for websites, games, design tools, animation, and other projects where [music] it needs to see whether its own work actually looks right. And since we're already talking about how businesses are using AI, here's something useful. Most of you voted for seven AI agents businesses are paying $3,000 to $10,000 for right now in yesterday's poll. So, we're giving it away for free. It breaks down the exact agents [music] businesses are buying, what they cost, who needs them, and how to spot the right clients. It also covers the ROI, setup fees, retainers, [music] and what a realistic $10,000 per month client list could look like with zero coding required. You can grab it through the link in the video description. Now, back to Kimmy K3. In one demonstration, K3 built a 3D open world game inside a browser using 3.js, JS WebGPU, and GPU [music] compute. It generated the environment and used outside tools to create a rider and horse. Moonshot also showed it building a simulation of China's Long March 10 rocket [music] and a Game Boy Advance emulator. And the company says K3 spent 15 hours improving GPU code and cut the required compute time by more than half. [music] It created a small GPU compiler called Mini Triton from scratch and reportedly designed a working chip over 48 hours using open-source engineering tools. Moonshot also says K3 reproduced an astrophysics analysis in about 2 hours. It reviewed more than 20 papers and wrote over 3,000 lines of Python code. According to the company, an experienced team might normally spend between 1 and 2 weeks on similar work. Now, these are company demonstrations, so they still need wider testing, but they show the goal clearly. K3 is supposed to stay on a difficult task for an entire afternoon, an entire night, or even several days. It can process up to 1 million tokens in one session, roughly 750,000 words. That means it can take in huge code bases, long documents, research papers, or project histories without immediately losing track of what came earlier. It also handles text, [music] images, and video inside the same model. Moonshot says K3 even edited its own promotional video from 56 clips. Its Kimmy work platform is adding interactive widgets and dashboards so users can build persistent workspaces [music] rather than using the model only through a normal chat box. The model uses a mixture of experts design and the simple explanation is that it contains hundreds of specialized sections, but it doesn't activate all of them for every question. K3 has 896 experts with only 16 active at a time. That lets Moonshot build a 2.8 trillion parameter system without using the entire model for every word it generates. Moonshot also created Kimmy Delta attention and attention residuals, two systems designed to help K3 [music] keep track of information through very long tasks. Combined with new training methods, the company claims they make K3 about 2.5 times more efficient at scaling than Kimmy K2. Now, the model was trained with MXFP4 weights and MXFP8 [music] activations, lower precision formats intended to reduce the hardware burden. That doesn't mean anyone will be running it on a normal laptop. And Moonshot recommends systems with at least 64 AI accelerators. The local LLaMA community immediately started joking that all you need is 2 terabytes of VRAM, several Mac Studios, a pile of storage drives, and enormous patience. [music] So, that's the strange reality of K3. It's open, but it isn't small. Most people will never host it at home. The companies that can host it, however, [music] are exactly the ones that matter. Large firms already spend millions every month on AI from OpenAI and Anthropic. If they can run K3 on their own servers, customize it, and keep their data private, [music] this becomes a serious threat to closed AI providers, and the price makes that threat stronger. K3 costs $3 per million input tokens when the input isn't cached, 30 cents when it's cached, and $15 per million output [music] tokens, including reasoning. Those prices stay the same even with long context. Fable 5 costs around $10 per million input tokens and $50 per million output tokens. GPT 5.6 Soul costs about 50 cents per million input tokens and $30 per million output tokens. So K3 is around

stood on a stage in Shanghai and told the world that China doesn't plan to follow America's rules for artificial intelligence. It wants to help write the rules, spread its own models across the developing world, [music] and build a

new AI order with China at the center. And Kim K3 gives that speech real weight because this isn't another model that looks impressive only inside a company presentation. Now, K3 has 2.8 trillion parameters, making it the largest openweight AI model ever announced.

Moonshot says it's the first open model to approach the 3 trillion mark and its full weights are expected to be released on July 27th. Companies and researchers will then be able to download it, modify it, and run it on their own

infrastructure instead of being permanently locked inside Moonshot's app. But the size isn't why this release is sending shock waves through the industry. And the real reason is that Kimmy K3 appears to be dangerously close to the best American closed models. And

in some tasks, it's already beating them. Moonshot says K3 can compete [music] with Anthropics Fable 5 and outperform Claude Opus 4.8, GPT 5.5, and parts of GPT 5.6 in demanding coding work. And independent testing actually tells a similar story. Artificial

analysis gave K3 a score of 57 on its intelligence index. Claude Opus 4.8 scored around 56, GPT 5.6, Six, Terra scored 55 and Gemini 3.1 Pro landed at about the same level as K3. So only Claude Fable 5 and GPT 5.6 Soul remained

ahead and even there the gap [music] was only two or three points. For years, the standard belief was that open models were 6 to 12 months behind the best closed systems. More recently, people started saying Chinese labs were perhaps

3 to 6 months behind. One Reddit commenter described the new reality more bluntly. Maybe they're six days behind.

Now, the most dramatic result came from arena.i's front-end code arena, where developers ask models to build real websites and interfaces, then vote on which result is better. Kimmy K3 entered at number one with 1,679 points. Claude Fable 5 scored 1,631.

GPT 5.6. Saul scored 1,618. K3 finished first in six of the seven areas tested, [music] including brand and marketing work and data dashboards.

Moonshot's previous model, Kimmy K 2.6, had been sitting in 18th place. K3 jumped 17 positions in one generation and went straight to the top. Even Arena's CEO called it potentially the biggest AI release of the year and suggested it could mark the moment China

moved ahead of the United States in at least one major part of the model race.

Now, Moonshot isn't pretending K3 wins everything. The company admits it still trails Fable 5 and GPT 5.6 Saul in overall user experience [music] and some broader tasks. But K3 doesn't need to dominate every category to be disruptive. It only needs to make

companies ask an uncomfortable question. Why keep [music] paying premium prices to American providers if an open alternative is almost as capable? So, Moonshot built K3 mainly for longrunning work, especially software development.

And this isn't a model designed only to answer one prompt and stop. It's meant to inspect a large codebase, create a plan, use tools, [music] make changes, check whether they worked, and continue for hours with limited human help. It

can also look at what appears on the screen. Moonshot calls this vision in the loop. The model writes code, looks at the result, notices what's wrong, changes the code, and checks again. That makes it useful for websites, games,

design tools, animation, and other projects where [music] it needs to see whether its own work actually looks right. And since we're already talking about how businesses are using AI, here's something useful. Most of you voted for seven AI agents businesses are

paying $3,000 to $10,000 for right now in yesterday's poll. So, we're giving it away for free. It breaks down the exact agents [music] businesses are buying, what they cost, who needs them, and how to spot the right clients. It also

covers the ROI, setup fees, retainers, [music] and what a realistic $10,000 per month client list could look like with zero coding required. You can grab it through the link in the video description. Now, back to Kimmy K3. In

one demonstration, K3 built a 3D openw world game inside a browser using 3.js, JS WebGPU and GPU [music] compute. It generated the environment and used outside tools to create a rider and horse. Moonshot also showed it building a simulation of China's Long March 10

rocket [music] and a Game Boy Advance emulator. And the company says K3 spent 15 hours improving GPU code and cut the required compute time by more than half.

[music] It created a small GPU compiler called Mini Triton from scratch and reportedly designed a working chip over 48 hours using open-source engineering tools.

Moonshot also says K3 reproduced an astrophysics analysis in about 2 hours. It reviewed more than 20 papers and wrote over 3,000 lines of Python code.

According to the company, an experienced team might normally spend between 1 and 2 weeks on similar work. Now, these are company demonstrations, so they still need wider testing, but they show the goal clearly. K3 is supposed to stay on

a difficult task for an entire afternoon, an entire night, or even several days. It can process up to 1 million tokens in one session, roughly 750,000 words. That means it can take in huge code bases, long documents, research papers, or project histories

without immediately losing track of what came earlier. It also handles text, [music] images, and video inside the same model.

Moonshot says K3 even edited its own promotional video from 56 clips. Its Kimmy work platform is adding interactive widgets and dashboards so users can build persistent workspaces [music] rather than using the model only through a normal chat box. The model

uses a mixture of experts design and the simple explanation is that it contains hundreds of specialized sections, but it doesn't activate all of them for every question. K3 has 896 experts with only 16 active at a time. That lets Moonshot

build a 2.8 trillion parameter system without using the entire model for every word it generates. Moonshot also created Kimmy Delta attention and attention residuals, two systems designed to help K3 [music] keep track of information through very long tasks. Combined with

new training methods, the company claims they make K3 about 2.5 times more efficient at scaling than Kimmy K2. Now, the model was trained with MXFP4 weights and MXFP8 [music] activations, lower precision formats intended to reduce the hardware burden.

That doesn't mean anyone will be running it on a normal laptop. And Moonshot recommends systems with at least 64 AI accelerators. The local llama community immediately started joking that all you need is 2 terb of VRAM, several Mac

Studios, a pile of storage drives, and enormous patience. [music] So, that's the strange reality of K3. It's open, but it isn't small. Most people will never host it at home. The companies that can host it, however, [music] are exactly the ones that matter. Large

firms already spend millions every month on AI from OpenAI and Anthropic. If they can run K3 on their own servers, customize it, and keep their data private, [music] this becomes a serious threat to closed AI providers, and the

price makes that threat stronger. K3 costs $3 per million input tokens when the input isn't cached, 30 cents when it's cached, and $15 per million output [music] tokens, including reasoning.

Those prices stay the same even with long context. Fable 5 costs around $10 per million input tokens and $50 per million output tokens. GPT 5.6 Soul costs about 50 cents per million input tokens and $30 per million output

tokens. So K3 is around five times more expensive than some earlier Kimmy models, [music] but it's still aggressively priced against Frontier Western systems. Moonshot says it launches [music] at maximum thinking effort by default with cheaper modes coming later. The company also admits K3

has weaknesses. If an agent system fails to return its full reasoning history, performance can drop. And when instructions are vague, K3 may make decisions on its own. So users who need strict control must set very clear rules. And even with those limits, the

release immediately reminded people [music] of DeepSeek R1. When Deepseek showed that a Chinese lab could match far more expensive American models, roughly $1 trillion was wiped from major technology stocks during [music] the panic that followed. Washington grew more concerned and the Trump

administration pushed even harder on technology export restrictions. K3 hasn't triggered that kind of market collapse, at least not yet. But it hit Chinese competitors almost [music] immediately. JIEPU shares fell 21.9% in Hong Kong while Miniax dropped [music] 13.8%.

So investors clearly understood what had happened. A new model had arrived with enormous scale, frontier level performance, open weights on the way, and pricing that could pressure almost everyone else. And Miniax is reportedly preparing its own 2.7 trillion parameter

model for release as early as the third quarter of 2026 along with a Frontier multimodal model called H3. Before K3, Mtoan's Longat 2.0 and Deepseek V4 Pro were among China's largest systems at around 1.6 trillion parameters. Several Chinese labs have now crossed the 1

trillion mark. But K3 isn't an isolated event. Z.AI's AI's GLM 5.2 recently shocked analysts by getting close to the best American closed models. Deepseek is still advancing. Miniaax is preparing larger systems and Chinese releases are becoming faster, cheaper, and more

capable. And Moonshot itself is backed by Alibaba and Tencent. Bloomberg reported that it's trying to raise $2 billion [music] at a valuation of around $30 billion ahead of a possible Hong Kong listing. There's also a political fight building around how these models

are trained. Anthropic previously accused Moonshot, Deepseek, and Miniax of using model distillation to copy capabilities from Claude in violation of its rules. Distillation is a common technique where one model helps train another. But US officials have started describing some forms of it as an

adversarial tactic. [music] And critics pointed out the irony. American AI companies trained their own systems on huge parts of the public internet, then became angry when other companies learned from their models. So, expect that argument to intensify when K3's

weights are released. There will be accusations about copied capabilities, scraped data, [music] export controls, and national security.

But those arguments may matter less if Chinese models keep improving this quickly. And that's where Xiinping's speech enters the story. At the World Artificial Intelligence Conference in Shanghai, Shei presented China as the leader of a new global AI order. He

called open-source AI a rare and historic opportunity and warned that unequal access could create new historical injustices. He compared AI to the invention of the steam engine and electricity, then offered developing countries something very different from the American model. lowercost open

technology, Chinese training, Chinese expertise, and a larger role in deciding how AI is governed. She promoted the new World AI Cooperation Organization or WICO, which signed up 29 countries the [music] day before his speech. He called it a milestone in AI history. China also

plans to build cooperation centers and training programs with BRICS, Azion, Latin America, and the African [music] Union. So this is a direct challenge to the US-led PAX silica initiative which is trying to secure AI infrastructure, semiconductor supply chains, and

critical minerals among American partners. She didn't name the United States, but he didn't need to. His message was that China won't accept American control over AI standards, advanced chips, global supply chains, [music] or access to powerful models. He

also made his strongest comments yet on AI safety. She said AI must remain under human control and called for early warning systems, emergency plans, and protection [music] against autonomous systems escaping human oversight. China is therefore presenting itself as the

country offering open access to the world while also claiming it can lead on safety and global standards. The World AI conference runs from July 17th to [music] July 20th. Attendees include major Chinese technology companies, UN Secretary General Antonio Gutirez,

Kazakhstan's President Kasim Jomar Tokayv [music] and Thailand's Prime Minister Anutin Charvira. It also comes just before the first government level AI talks between China and the United States under President Donald Trump. At a UN AI meeting last week, American officials

argued that too much regulation could slow innovation. China pushed its own message. lowcost open models could reduce the global technology gap. So now China has Kim K3 to point to. It's no longer promising that its open AI ecosystem might become competitive one

day. It's released a model that's already beating some of America's biggest names, costs less to use, and will soon be available for companies to run themselves. All right, that's it for this one. Let me know what you think in

the comments. Thanks for watching, and I'll catch you in the next one.