- AI Valley
- Posts
- Alibaba unveils advanced Qwen 3 AI
Alibaba unveils advanced Qwen 3 AI
PLUS: Viral Rumors of DeepSeek R2 Leaked
Together with
Howdy again. It’s Barsee, and welcome back to AI Valley.
Today’s climb through the Valley reveals:
🚨 Viral Rumors of DeepSeek R2 Leaked
🛍️ OpenAI Turns ChatGPT Into Shopping Assistant With New Search Update
💡 Alibaba unveils advanced Qwen 3 AI
🤖 Plus trending AI tools, posts, and resources
Let’s dive into the Valley of AI…
Leading companies are 38% more likely to use an AI-native email app — they don't just settle for Gmail or Outlook.
Superhuman is trusted by the world's fastest-growing companies, including over 60% of the Forbes AI 50.
Fly through your inbox with AI that helps you organize messages, schedule calls, write emails, remember follow-ups, and more.
Emails that write themselves. When it's time to follow up, Superhuman will remind you and draft the email — all you have to do is send it.
A more focused inbox. Superhuman can automatically archive cold pitches, marketing emails, and more via Auto Labels. Create your own labels with AI prompts. Eliminate inbox clutter for good.
AI that sounds like you. Superhuman AI adapts its writing style to the person you're emailing and to your unique voice and tone.
*This is sponsored
THROUGH THE VALLEY
🚨Viral rumors of DeepSeek R2 leaked!
—1.2T param, 78B active, hybrid MoE
—97.3% cheaper than GPT 4o ($0.07/M in, $0.27/M out)
—5.2PB training data. 89.7% on C-Eval2.0
—Better vision. 92.4% on COCO
—82% utilization in Huawei Ascend 910BBig shift away from US supply chain.
— Deedy (@deedydas)
4:00 PM • Apr 26, 2025
Leaked reports on Chinese finance forums suggest DeepSeek’s upcoming R2 model packs 1.2 trillion parameters, runs on Huawei’s Ascend 910B chips, and rivals GPT-4.5, while costing 97% less. Though DeepSeek denies the leaks, if true, this could shake up the AI industry, slashing costs and challenging Western dominance.
OpenAI is rolling out a shopping-friendly search mode in ChatGPT, letting users compare prices, read reviews, and buy products directly in the chat. Results include images, pricing, and purchase links—no ads or paid placements. The feature will soon be available to all users, including free tiers.
Hugging Face just launched the SO-101, a $100 programmable robotic arm you can 3D-print at home. Built with partners like The Robot Studio and WowRobo, it’s faster to assemble than its predecessor and has better motors and AI training. Plus, after acquiring Pollen Robotics, they’ll soon sell Reachy 2, a humanoid bot led by an ex-Tesla engineer.
Following Shopify’s lead, Duolingo is replacing contractors with AI for tasks like content creation. CEO Luis von Ahn says AI will handle repetitive work, while humans focus on creative problem-solving. New hires? Only if AI can’t do the job. AI is also being used in hiring and performance reviews.
We got a robot to clean up homes that were never seen in its training data! Our new model, π-0.5, aims to tackle open-world generalization.
We took our robot into homes that were not in the training data and asked it to clean kitchens and bedrooms. More below⤵️
— Physical Intelligence (@physical_int)
4:56 PM • Apr 22, 2025
OpenAI-supported startup Physical Intelligence just unveiled π-0.5, a smarter robotics model that can generalize and plan for unseen tasks. In demos, it cleaned kitchens and bedrooms effortlessly. The team also open-sourced earlier models (π-0 & π0-FAST) on GitHub for developers.
Tavus just dropped Hummingbird-0, a zero-shot lip-sync model that matches lip movements to any voice—no training needed. Built from their Phoenix-3 model, it produces hyper-realistic sync from just one video input, outperforming all competitors. Test it now on Fal’s platform or via Tavus’ API.
PEAK OF THE DAY
💡 Alibaba unveils advanced Qwen 3 AI
Alibaba just released the Qwen 3 family of language models, ranging from 0.6 billion to 235 billion parameters. These models bring significant improvements in reasoning, coding, math, and multilingual tasks while staying efficient and fully open source.
What makes it special?
MoE (Mixture of Experts) Models: Two of the largest models, Qwen3-30B-A3B and Qwen3-23 5B-A22B, use MoE architecture, which activates only a portion of the model per query. This makes them more efficient and cheaper to deploy than models like DeepSeek V3.
Hybrid thinking mode: Qwen 3 intelligently switches between "thinking mode" for step-by-step reasoning and "non-thinking mode" for quick responses to simpler questions.
Massive training data: Trained on a whopping 36 trillion tokens, double the size of Qwen2.5's training data. This helps it reason more effectively, follow complex instructions, and handle tool use better.
Multilingual Mastery: Supports 119 languages and dialects, making it highly effective at translation and following multilingual instructions for a wide range of global applications.
Advanced agent capabilities: The models natively support Model Context Protocol (MCP) and advanced function calling. This makes Qwen 3 ready for real-world agent tasks like managing workflows or controlling devices such as PCs or smartphones.
How well does it perform?
It's not dominant across every domain, but it's highly competitive, especially for programming and math tasks. The flagship Qwen3-235B-A22B model outperforms OpenAI's o3-mini and Google's Gemini 2.5 Pro on benchmarks like Codeforces and AIME25.
Where can you get it?
All the Qwen3 models are open-sourced under the Apache 2.0 license and are now available on Hugging Face, GitHub, and ModelScope.
Why does it matter?
Qwen 3 blends performance, flexibility, and openness in a way that challenges both domestic rivals like DeepSeek and global players like OpenAI and Google. With its hybrid reasoning system and MoE efficiency, it offers enterprise-grade results without enterprise-grade costs.
TRENDING TOOLS
Lambda > Run DeepSeek models for less. V3: $0.34/M input tokens, $0.88/M output tokens. R1: $0.54/M input tokens, $2.18/M output tokens. *
Keytake > AI video assistant that helps you turn ideas, documents, or web content into professional-grade videos in minutes.
PageOn 2.0 > Cursor for visual communication, beyond slides.
Cluely > Invisible AI assistant that helps you cheat smartly during virtual interactions like sales calls and online meetings.
NeuralAgent > The first AI agent that lives and works on your desktop.
Peek > AI personal finance coach that guides you through decisions.
Agent Simulate > Test your AI apps with thousands of digital humans.
Corgea > Ship fast without worrying about security.
THINK PIECES / BRAIN BOOST
Can OpenAI win at search?
Create an AI UGC farm using arcads.ai by Jeddi.
I trained AI to write 30 viral tweets (Build an audience faster) by Dan Koe.
This is a prompting masterclass for vibe marketing.
AI's impact on software development, a research by Anthropic.
AI can handle tasks twice as complex every few months.
Google's new guide on building multi-agent apps.
How to share AI’s future wealth.
Teens are using ChatGPT to invest in the stock market.
VALLEY GEMS
BREAKING: The University of Zurich has been using AI bots to secretly manipulate Redditors since November 2024.
The scariest part?
The bots were 6 times more likely to change the minds of Redditors than the baseline, often by leveraging misinformation.More below🧵
— Reddit Lies (@reddit_lies)
6:03 PM • Apr 28, 2025
Prompts that changed my life. Part 9
— Prompter (@PromptLLM)
8:31 PM • Apr 28, 2025
This is the ultimate guide to vibe coding. Save it.
— Aadit Sheth (@aaditsh)
2:04 PM • Apr 27, 2025
THAT’S ALL FOR TODAY
Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee
👍️ New reader? Subscribe here
Thanks for being here.
HOW WAS TODAY'S NEWSLETTER |
REACH 100K+ READERS
Acquire new customers and drive revenue by partnering with us
Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.
If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.