- AI Fire
- Posts
- 🇨🇳 China’s Open AI Lead Widens
In partnership with
China’s open-model race is getting wild. Kimi K3 hit 2.8T parameters, while most U.S. releases stayed far smaller. Hugging Face’s new data shows a pretty dramatic shift!
What's on FIRE 🔥
IN PARTNERSHIP WITH SECTION
On 9/17 from 11am–6pm ET, join Scott Galloway, Section and AI leaders from Wayfair, MetLife, TD Bank, Booz Allen, and more for the bi-annual AI:ROI Conference.
One day. Ten sessions. You’ll walk away with a methodology for AI ROI rooted in experience and reality.
AI INSIGHTS
Hugging Face’s latest analysis just shows that Chinese labs are now releasing much larger frontier open models than U.S. labs.
From January through August 2026, the largest high-performing open model released by a Chinese lab was bigger than the largest U.S. release in almost every month. Some of the biggest examples include:
Moonshot’s Kimi K3 at roughly 2.8T parameters
Large models from Xiaomi, Ant Group, and Meituan that also crossed the 1T parameter mark
China’s monthly top-end releases ranging from roughly 754B to 2.78T parameters
U.S. releases staying below 130B parameters in five of the first 7 months
Labs like Moonshot, MiniMax, Xiaomi, and Z.ai focus heavily on large frontier models, while Qwen and Tencent release models across many sizes.
The U.S. still leads in overall open-model activity, with NVIDIA and AMD publishing hundreds of repositories. But at the frontier end, China is moving faster. However, when you see frontier-scale open weights, China is becoming much more aggressive.
PRESENTED BY HUBSPOT
Winning Startups Aren't Bigger. They're Leaner.
Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.
Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.
AI SOURCES FROM AI FIRE
1. How to Manage AI Agents Like a Real Team in 2026. Learn how to run 5–10+ AI agents at once, delegate work, build automation loops, control costs, and turn agents into a working team you can actually manage.
2. I Tested Grok 4.6 on Real Coding Projects. Here’s How Good It Actually Is. See how Grok 4.6 handles games, full websites, bug hunting, and long AI agent workflows, plus where its lower pricing makes it worth using.
3. Video: 3+ Ways To Give AI Perfect Memory. ChatGPT, Claude, and Gemini can remember your preferences, but still forget important context. Learn how to build a simple AI memory system using personal, project, and portable memory workflows.
FREE LIVE WEBINAR THIS WEEK
Most people are only using AI at 5-10% of what it can actually do. The difference between an amateur and a top AI builder is simple: instead of using AI to do one job, they design a system that dictates how a hundred jobs get done at once.
This week, we’re running a 100% FREE live workshop to show you how to build your first multi-agent workflow from scratch using the Graph Engineering technique:
🧠 Understand "Graph Engineering" in simple terms.
✂️ Split complex tasks across AI agents to do the heavy lifting.
🤝 Make AI agents run in parallel to fact-check and combine results.
💬 Join a live Q&A to apply this workflow to your specific needs.
📅 Wednesday, August 19
⏰ 10:00 PM ET / 7:00 PM PT
🎟️ Free to join
No coding needed. By the end of this session, you’ll understand the exact logic to turn scattered ideas into a smooth, self-running machine of AI agents working together on autopilot.
Can’t make it live? Register anyway and we’ll send you the replay.
TODAY IN AI
AI HIGHLIGHTS
💻 Cursor just launched Origin, a new code hosting platform with repos, pull requests, GitHub sync, and Cursor agents built in. It’s rolling out in early beta for paid users.
😂 This is the funniest ‘AI startup’ today. SF residents thought they were chatting with ChatTJB, but some guy named Tucker was replying for real. Here's this wild story.
🎧 Pika Labs launched 4 audio models for video soundtracks, music, sound effects & voice. Some are dramatically cheaper than competing models. So try using here.
🎬 ByteDance reached Hollywood’s first major AI copyright agreement with the Motion Picture Association. New protections will cover Seedance and Seedream across apps like TikTok, CapCut, and Dreamina.
⚡ OpenAI is planning a massive 8 GW AI data center in Ohio with NVIDIA and SB Energy. The project could create 35,000 construction jobs, with the first 800 MW expected in 2028.
💰 Big AI Fundraising: Wispr raised $280M, bringing total funding to $361M. The company is expanding beyond dictation, while its new Canto model cuts speech error rates from 30% to under 10%.
NEW EMPOWERED AI TOOLS
🤖 Omni by Xpander takes AI agents from your laptop to the cloud, making them scheduled, long-running, shareable, and maintained.
📝 Meridian automatically tracks your work and turns it into daily summaries and ready-to-post Jira updates, all locally, free, and open source.
🤖 Clears lets autonomous agents handle software delivery across the entire SDLC, helping R&D teams ship much faster.
🧩 Vendo lets you build your own features and micro-apps inside your product just by simply describing what you need.
AI BREAKTHROUGH
A Beijing neurosurgery resident, Shanmu Jin, reportedly used GPT-5.6 Sol in ChatGPT Work to help solve Crouzeix’s Conjecture, a matrix-analysis problem that had remained open since 2004.
FYI, he’s not even a professional mathematician. He originally studied geology, later trained in medicine, and reached the problem through research related to brain ultrasound. His setup was also more involved than simply asking ChatGPT for a proof.
Jin reportedly ran GPT-5.6 Sol for around 16 hours in an autonomous session.
The model had no internet access, so it could not simply search for an existing solution.
Jin used multiple subagents that challenged one another, looked for flaws, and tried competing approaches.
The workflow was reportedly inspired by one of OpenAI’s own math prompting methods.
GPT-5.6 Sol eventually surfaced a proof that Jin then examined and developed further.
The result has not completed formal peer review yet, but several respected mathematicians have reportedly checked the argument. These include Cornell mathematician Alex Townsend, Anne Greenbaum, and Michel Crouzeix.
Every step has to survive expert scrutiny, especially for a problem that resisted researchers for more than 2 decades. We’ll wait and see!
Blu Dot surpasses 2,000% ROAS with self-serve CTV ads
Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:
After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.
The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.
“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”
Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply