Tech Brew Ride Home - Kimi K3

Episode Date: July 17, 2026

Moonshot AI released Kimi K3, a 2.8T-parameter model it says rivals Opus 4.8 and GPT-5.5. Google fell months behind on Gemini 3.5 Pro, MLB banned dugout iPads from accessing GenAI for in-game calls, and The Verge tested Siri AI. Moonshot AI releases Kimi K3, a 2.8T-parameter AI model that it says rivals Claude Opus 4.8 and GPT-5.5, and plans to release its full model weights by July 27 (VentureBeat) Sources: Google is months behind schedule on delivering Gemini 3.5 Pro as it tries to improve its capabilities, particularly in coding; GOOG closes down 4.43% (Bloomberg) Memo: MLB bans the use of league-provided dugout iPads to access GenAI for in-game strategy calls; sources say at least a third of teams used AI this way (The Athletic) Longreads The Verge spends a month testing Siri AI in the iOS 27 public beta, finding it's already reshaping how people use their iPhone, though it can't yet reach non-Apple apps (The Verge) Subscribe to the ad-free feed. Learn more about your ad choices. Visit megaphone.fm/adchoices

The Hulu original series Furious is coming to Disney Plus. Starring Emmy Rossum, Furious follows FBI agent Alice Black on the hunt for a mysterious and calculating serial killer. Both walk their own paths toward justice, and as their lives start to intertwine, the line between right and wrong begins to blur. Don't miss the three-episode premiere of the Hulu original series Furious on July 27th, only on Hulu on Disney Plus. Welcome to the Tech Brew Ride Home for Friday, July 17th, 2026. I'm Brian McCullough today. Moonshot AI released Kimmy K3, a new model that they say rivals Opus 4.8 and GBT 5.5. Google has fallen months behind on its Gemini 3.5 Pro, Major League Baseball Band dugout iPads from accessing AI in game.

And of course, the weekend long read suggestions. Here's what you missed today in the world of tech. To all the IT pros out there, today's sponsor has a special gift for you. That gift is Ace of Uptime, an online card. based game where you go up against the problems that threaten uptime on a daily basis. Each level pits you against a recognizable adversary. We're talking about alert overloads, heat that's trapped in a cramped server room, and systems that look fine but are far from it. Choose your move and see whether your decisions resolve the issue or escalate it. The game is fast, fun, and designed to let you test how you'd respond when uptime is on the line. Think you can beat the villains of downtime. Head over to eton.com slash ace to find out. That's eaton.com slash ace.

A.I. Company, Moonshot AI, has released Kimmy K3, a 2.8 trillion parameter AI model that it says rivals Claude Opus 4.8 and GPT5.5 and plans to release its full model weights by July 27th. Kimmy K3 has taken the number one spot on the front-end Code Arena benchmark, surpassing Claude Fable 5 and scoring 88.3 on Terminal Bench 2.1, which trails only GPT 5.6 Sol's score of 88.8. Quoting Venture Beat. If you want to take Kimmy K3 for a spin right now, you can just head to Kimmy.com, sign up with a Google account or phone number, no credit card required, and start chatting with what may be the most powerful open source model ever built. Kimmy K3 is a frontier class large language model with 2.8 trillion total parameters, roughly 75% larger than Deepseek's

V4 Pro, which the company's own timeline chart shows at approximately 1.6 trillion parameters. The The model features a 1 million token context window, native visual understanding capabilities, and an always-on reasoning mode that the company calls thinking mode. The model is built on two key architectural innovations developed internally at Moonshot AI, Kimi Delta Attention, a hybrid linear attention mechanism and attention residuals, which the company describes as a drop-in replacement for residual connections that deliver consistent scaling gains. Both techniques were previously published as open research by the Moonshot team on

GitHub. On the API side, Kimmy K3 is compatible with the OpenAI SDK, lowering the integration barrier for developers already building on OpenAI or Anthropic Tool Chains. The model is priced at $3 per million input tokens and $15 per million output tokens with cashed input tokens dropping to just 30 cents per million, pricing that positions it roughly in line with mid-tier offerings from Western Labs. But at a performance level, the company claims approaches the top of the market. A promotional top-up rebate, running through August 12 offers up to 30% back in vouchers for API credits of $1,000 or more. The benchmark results drawn from public leaderboard data and a private evaluation by

analytics firm artificial analysis tells a striking story. On GDP Val AAV2, a benchmark measuring real-world tasks across 44 occupations and nine major industries, Kimmy K-3 scored 1,6807, placing it third overall behind only Claude Fable 5 Max at 1815, and GPD 5.6 sole max at 1747.8 and ahead of Claude Opus 4.8 at 1600. On AA briefcase, a private agenetic benchmark from artificial analysis designed to test long horizon knowledge work, K3 climbed to second place, beating GPD 5.6 sole max and trailing only Fable 5 max. Perhaps most impressively, K3 achieved a state-of-the-art score of 91.2 out of 100 on Browse Comp, a benchmark for Long Horizon high-difficulty information.

seeking. The company says it accomplished this in a single agent setup using its 1 million token context window without any context compression or additional context management techniques. A feat that suggests raw context length when paired with strong retrieval capabilities may be more powerful than elaborate multi-agent workarounds. As one widely followed AI commentator put it on social media, open source is no longer lagging six months behind Western closed source models. Read that again and think about what it all means. That observation captures the significance of the moment for much of the past three years, open source models have typically trailed their proprietary counterparts by a meaningful margin.

Kimmy K3 appears to have closed that gap almost entirely. Beyond Raw Benchmarks, Moonshot AI, showcased a proof of concept that may be even more revealing of K3's capabilities and the company's strategic direction. In a demonstration documented in the company's technical materials, Kimmy K3 was tasked with designing a physical chip to run a nanoscale version of itself. Over 48 hours, of continuous autonomous agent operation. K3 independently completed the chip's full construction pipeline from architectural design through optimization and verification using open source electric design automation tools. The result was a tiny but functional chip design just four square millimeters that achieved timing convergence at 100 megahertz and could decode more than

8,700 tokens per second in simulation. This is not a production chip. It is a demonstration of what Moonshot AI clearly views as the next competitive frontier, long-range autonomous agent capabilities. The ability to sustain coherent, multi-step technical work over a 48-hour window, reading documentation, making design decisions, running verification loops, and iterating on failures represents a qualitative leap beyond the kind of single-turn question and answering that defined the first generation of large language models. The company also highlighted a case in computational astrophysics where K3 reportedly reproduced the universal I Love Q relation, a complex calculation that typically takes a senior researcher one to two weeks in approximately two hours,

reading and cross-validating more than 20 papers and implementing a complete numerical pipeline along the way. For enterprise technology leaders, the implications are concrete. A 2.8 trillion parameter open source model that performs at near-frontier levels creates new options for companies that want to fine-tune, self-host, or build proprietary systems on top of a capable base model without being locked into API contracts with OpenAI or Anthropic. The trade-off, of course, is that running a model of the size requires substantial GPU infrastructure. Inference at 2.8 trillion parameters is not something that runs on a single server rack. The performance gap between open source and proprietary models has functionally closed at the frontier.

If K3's benchmark numbers hold up under independent evaluation, and particularly once the open weights are available for community testing on July 27th, it will be difficult for closed source providers to justify premium pricing purely on the basis of capability. The locus of AI innovation, meanwhile, continues to shift. China's AI ecosystem, which many Western observers questioned after early struggles with chip export restrictions, has now produced a model that competes with the best systems from companies with direct access to Nvidia's most advanced hardware. The architectural innovations behind K3, particularly the hybrid linear attention mechanism, suggests that algorithmic efficiency may matter as much as raw compute.

And the agenic capabilities demonstrated by K3, chip design, multi-week research compression, long horizon information seeking, point toward a future where AI models are not just answering questions, but autonomously executing complex multi-day projects. For enterprises evaluating AI investments, this shifts the value proposition from productivity co-pilot to autonomous technical workforce, end quote. Meanwhile, you know who we haven't heard from in a while? Quoting Bloomberg, Alphabet's Google is months behind schedule on delivering Gemini 3.5 its most powerful flagship AI model because the company has been taking time to try to improve its capabilities, particularly in coding, according to people familiar with the matter.

The delay has been a source of frustration for Google engineers, AI researchers, and managers, many of whom are concerned the company risks losing an edge in the market as rivals Anthropic and Open AI produce models that exceed Gemini's capabilities, according to 10 current and former employees. Google has multiple layers of stakeholders involved in preparing models for release, working to weave AI across a vast product portfolio, including search, maps, and YouTube, which can cause delays, said the people who declined to be named discussing internal concerns. Both OpenAI and Meta recently released new models that further outpaced Google's current offerings in AI for writing code. Late last month, Google updated the data being used to train Gemini in an attempt to improve these skills, but the results were disappointing, one of the people said. Google's popular products are a gateway degenerative AI for everyday people and can yield data that makes their answer smarter, but encouraging leadership of every department to move in the same direction is like trying to boil an ocean one,

ex-employee said. When mandates shift or efforts end up duplicated in multiple departments, it gets even more difficult to maintain a cohesive strategy, current and former employees said. It's also a challenge for anyone offering to get the resources it would need to succeed and to gain traction in the market, they said. Google co-founder Sergei Bryn and others were advocating for Google to move faster to seize opportunities in AI coding, but their efforts were slowed by competing factions within the company to former employees said. Cloud Computing Unit Google Cloud, Research Lab, Google DeepMind, and the team behind the Android operating system are all building AI coding tools for developers with involvement from some consumer

product teams to people familiar with the work said. Efforts to win at coding have also been up against some engineers at Google with a more purest stance who believe that all important codes should be human-written to adhere to Google standards, ex-employees said. Early in the rollout of the technology, employees also faced for restrictions on using Gemini to write or analyze software over concerns that proprietary code could leak into the AI models training data, they said, those policies which have since been relaxed, limited opportunities for engineers to experiment with AI development, end quote.

This episode is brought to you by Accenture. When your advertising operations fall out of sync, everything else follows. Spotify and Accenture are working together to reinvent the rhythm of ad sales, using automation, analytics, and smarter workflows to simplify campaign delivery and access better data across the business. The result? Less time spent on operations, more time connecting brands with the moments and fandoms that matter most.

Learn more at Accenture.com slash Spotify. There's so much you need to remember to say during your big meeting, but you're human. And sometimes in the heat of the moment, a few points can slip your mind. That's why even G2 exists. Their productivity smart glasses designed with teleprompting, conversation support,

real-time translation, and more. They're also made the look and feel like premium eyewear with no camera and a lightweight 36 gram design you can wear all day. Think of them as your sidekick, helping you stay on script and on point. To learn more about EvenG2, go to Evenrealities.com. Use promo code TechBrew to get 10% off Even Ring 1 and or Even Clip when you add them to your Even G2 order. Major League Baseball has banned the use of League provided dugout iPads to access generative AI for in-game strategy calls. Sources say at least a third of teams were already using AI this way. Quoting the Times, according to a commissioner's office memo obtained by the athletic,

teams were pushing the boundaries of guidelines governing the use of technology by, in many cases, installing custom apps that expanded the use of the iPads beyond their originally intended purpose to include recommendations regarding substitutions, pitch-calling and other in-game decisions traditionally made by players and coaches. As much as a third of the league used the dugout iPads for at least one of these purposes, according to people with knowledge of the technology, who spoke on the condition of anonymity. The league memo was issued June 11th and gave teams more than a month to adjust before the band took effect on Wednesday in time for the start of the season's second half Thursday. The mid-season policy change has been met with frustration by some front office members tasked with innovation.

It's caused quite the stir, said one high-ranking of, official of an MLB team's research and development department, but even members of more data-friendly front offices say the trend raised eyebrows within the industry because of how many decisions made by managers and coaches could be aided by or even potentially replaced by technology, such as pitch-calling or player substitutions. MLB seemed to harbor similar concerns. The league's review ultimately found all clubs were compliant with the league's rules governing sign-stealing and electronic device usage, per MLB's memo. The tech band, therefore, did not come with punishment.

However, the issue proved urgent enough for the league to act mid-season in hopes of curbing the growth of these apps. Said one front office executive got to stop the cheating before there's cheating now. MLB declined to comment but referred to the memo which was issued by the league's executive vice president of baseball operations Morgan Sword. The players union also declined to comment. League provided iPads were first introduced into dugouts league-wide in 2016, though their usage was more tightly regulated in 2020, after signed stealing scandals rocked the sport. For the first time, we're going to be allowing players to view live in-game video via an iPad in the dugout or bullpen during the game, former MLB, Executive Chris Maranak said in March 2021, the iPads are issued by MLB and fully controlled so that players can only use the app that we've put on the device.

They can't access the internet for browsing or social media or any other kind of functions. The iPads are completely locked down and monitored with the software that we have on there. Entering this season, the league issued iPads provided access to data through three tabs, person with knowledge of the system who spoke on the condition of anonymity. The first contained all the MLB provided stat-cast data and multiple video angles. The second tab contained all data related to the automated ball strike system. The third tab, known across the league as the custom tab, housed each team's specialized data akin to information found in old-school paper-filled binders.

Examples include matchup info, defensive positioning, and player tendencies, which teams argued were necessities. Going forward, teams can still upload statutes. information, which is data that would be available before first pitch, though clubs were told that anything uploaded would ultimately remain subject to review by MLB, end quote. Only one long read this weekend, and it's also a bit of a weekend project, too. Maybe you should download the new iOS public beta this weekend to try out the new A-I-ified Siri, because lots of people are saying it's already changing how they use their iPhone.

Quoting the verge, before you went to an app and told the app what you wanted to do. call a car, set a timer, order lunch. Now you say what you want to do first, and Siri AI tries to look through all the apps and information available to it to handle the rest. So when I asked about a concert, Siri looked at what was on the webpage, then searched the web, found the answer, and presented it to me. I didn't need to jump around browser tabs or look at the Bandon Questions Instagram page. It was just there. In the month, I've been using Siri AI. It surprised me in tons of different ways. On my first day testing the beta during Apple's dev conference, I was able to ask, can you add my WWDC briefings to my calendar?

And Siri looked in my email, parsed the data, and added six individual events with correct times to my calendar. I'll note it could only add it to my Apple calendar, but I'll get to why shortly. These integrations have genuinely altered my brain chemistry a bit. Now I almost always try to use Siri first just to see if it can perform the action I need or answer a simple question. It's practically stopped me from opening my browser for most things since it's easier, faster, and more enjoyable to just swipe down from the top of the screen and type a prompt. On-screen awareness has probably been the most helpful addition for me. Being able to ask Siri about what's on my screen saves me a lot of tapping around. When it's able to take action from that on-screen awareness, like adding an event to my calendar

or directing me to an address on my screen, it's even better. More often than I've expected, Siri is able to do what I ask it to, and when it performs a somewhat complex task, it feels like magic. But when it hits a wall, I'm reminded of why it's going to take a bit of work to get it to the It Just Works universe, where Siri not working is the exception, not the norm. Right now, if you're on the Siri AI preview on the iOS 27 beta, Apple apps are the only ones with access to series new capabilities. If you live inside Apple's ecosystem, everything is great. Your data probably exists inside messages, mail, and photos. And when you want to take action, you're adding to-do items to reminders and notes. I use a few of these apps daily. And when they

work together, like adding a list of events from my email to my calendar, it really is like a glimpse at the future. But if I asked Siri, when did Daniel say he was free to play Dota? Then Siri won't have any idea because Daniel and I only message each other through telegram, which the system doesn't have access to, end quote. No weekend bonus episodes for you this weekend. Talk to you on Monday.