വീണ്ടും DeepSeek 🇨🇳?

We all thought the AI fight was simple.
Google vs OpenAI.
Two giants. One crown. 👑

But while they were busy punching each other in the face…
A new player broke the door, flipped the table, and said “Move.”

DeepSeek the open-source underdog from China just dropped V3.2-Speciale.
And the results? Honestly shocking.
It’s not just “good for a free model.”
It’s beating Google and OpenAI’s most expensive models in math and coding.
And it’s cheap. And open. And growing fast.

No wonder OpenAI is panicking.
Rumors say Sam Altman called a “Red Alert”, rushing a secret model called Garlic to stop the attack.
Is the king… actually scared? 👀

And for the creative people Alibaba just gave us a surprise:
A pro-level image model that runs on your home PC.

Giants are shaking.
Underdogs are rising.
Everything is changing again.

I’m Alex. Welcome to L8R by Innov8.
Let’s DIVE DEEP 🧠👇

In today’s post:

• OpenAI’s Secret “Garlic” Model – Can It Beat Gemini?

• Alibaba’s Z-Image – Open-Source Image Beast

• DeepSeek-V3.2 & V3.2-Speciale – Reasoning Monsters

🧄 OpenAI’s Secret “Garlic” Model – Can It Beat Gemini?

OpenAI is secretly training a new AI model called Garlic.
Early tests say it beats Google’s Gemini 3 and Anthropic’s Opus 4.5 in coding and logic.
The big idea: small model, big brain.

🔍 Key Points:

• Garlic is an inside code name for a new OpenAI model.

• It is built to be better at reasoning and coding than Gemini 3 and Opus 4.5.

• Main trick: a new pre-training method that packs more knowledge into a smaller model.

• This also fixes problems in GPT-4.5 and makes training cheaper and faster.

• Sam Altman called a company-wide “red alert” to focus on beating Google.

• Other side projects like ads and extra agents were paused to work on Garlic.

• Garlic may later come out to the public as GPT-5.2 or GPT-5.5.

📰 What’s New?

• Garlic uses an optimized training pipeline, so it can learn more from less data and less compute.

• Internal tests say it wins in programming and logic tests vs Gemini 3 and Opus 4.5.

• OpenAI is using Garlic as its main weapon in the new AI model war.

🚨 Why It Matters

• Shows OpenAI is worried about losing its lead to Google and Anthropic.

• If smaller models can act like big ones, more people and startups can use powerful AI.

• Better reasoning models mean stronger agents, better coding tools, and smarter ChatGPT-style apps.

⏭️ What’s Next?

• Garlic still needs post-training, safety checks, and more tests.

• Early version is expected sometime next year.

• It may later be branded as part of the GPT-5 family (like 5.2/5.5)
  

From our Partners

The Hustle: Claude Hacks For Marketers

Some people use Claude to write emails. Others use it to basically run their entire business while they play Wordle.

This isn't just ChatGPT's cooler cousin. It's the AI that's quietly revolutionizing how smart people work – writing entire business plans, planning marketing campaigns, and basically becoming the intern you never have to pay.

The Hustle's new guide shows you exactly how the AI-literate are leaving everyone else behind. Subscribe for instant access.

Get our Claude playbook

🖼 Alibaba’s Z-Image – Open-Source Image Beast

Alibaba launched Z-Image, an open-source image model.
It makes sharp, realistic images and even good text inside images, while running on normal GPUs at home.

🔍 Key Points:

• Z-Image is a 6B parameter text-to-image model.

• It uses a single stream for text and image tokens (S3-DiT), which makes it fast and efficient.

• There are three main versions:

  • Base – best quality

  • Turbo – faster, fewer steps

  • Edit – for changing existing images

• Trained with a smart data system that filters bad images and covers many real-world topics.

• Runs on consumer GPUs with <16 GB VRAM, so even hobby users can try it.

• Community says quality is as good as or better than SDXL, with great text rendering in Chinese and English.

📰 What’s New?

• Z-Image brings photorealistic images + clear text in a model that is fully open-source.

• Turbo can give results in very few steps, sometimes under a second on strong GPUs.

• The team shared the code + weights, so everyone can study and tweak it.

🚨 Why It Matters

• Proves that you don’t need a giant model to get top-tier image quality.

• Makes high-end AI art tools available to solo creators, students, and small studios.

• Strong text rendering means better posters, thumbnails, ads, UI mocks, etc.

• Gives the open-source world a serious rival to SDXL, Flux, and other big models.

⏭️ What’s Next?

• More LoRA fine-tunes, custom versions, and plug-ins for tools like ComfyUI will keep growing.

• Z-Image Edit and reward-based training may make editing and following instructions even better.

Together with guidde

You can (easily) launch a newsletter too

This newsletter you couldn’t wait to open? It runs on beehiiv — the absolute best platform for email newsletters.

Our editor makes your content look like Picasso in the inbox. Your website? Beautiful and ready to capture subscribers on day one.

And when it’s time to monetize, you don’t need to duct-tape a dozen tools together. Paid subscriptions, referrals, and a (super easy-to-use) global ad network — it’s all built in.

beehiiv isn’t just the best choice. It’s the only choice that makes sense.

Start free today. No credit card required

🧠 DeepSeek-V3.2 & V3.2-Speciale – Reasoning Monsters

DeepSeek’s V3.2 and V3.2-Speciale are new AI models focused on reasoning first, chatting later.
They are open-weight, cheap to use, and are already beating top closed models in math and coding tests.

🔍 Key Points:

• Built as a huge Mixture-of-Experts (MoE) model with smart sparse attention.

• DeepSeek Sparse Attention lets the model look only at the most important tokens, so it is cheaper and faster with long context (up to 128k tokens).

• V3.2-Speciale is a “thinking mode” model that reasons a lot before answering.

• On tough math tests like AIME and HMMT, Speciale scores higher than GPT-5 High and Gemini 3 Pro.

• On coding tasks like SWE-Bench, it solves a big chunk of real GitHub issues.

• Speciale is API-only (no app UI) because it uses a lot of tokens and compute.

• Pricing is very low per 1M tokens, and weights are under an MIT licence, allowing commercial use.

📰 What’s New?

• DeepSeek trained these models with agent-style tasks: tools, multi-step problems, planning, and chain-of-thought.

• The chat format includes a developer role that tells the model when and how to think and use tools.

• Speciale even won gold-medal level scores on math and programming olympiad benchmarks.

🚨 Why It Matters

• Shows that open models can now match or beat top closed models in reasoning and coding.

• Makes strong reasoning power affordable for startups, indie hackers, and research labs.

• Pushes AI closer to true agents that can plan, search, and solve tasks with less human hand-holding.

⏭️ What’s Next?

• DeepSeek plans to improve world knowledge, reduce token waste, and add better tool use (especially for Speciale).

• Future versions may extend context even more and support richer agent workflows and plugins.

🚨 Quick L8R Summary

• DeepSeek (an open-source rival) just dropped a new model that is literally beating OpenAI and Google’s best tools in math and coding.

• Rumors say OpenAI is in "Red Alert" mode, secretly training a powerful small model named "Garlic" to win the war against Gemini.

• Alibaba released Z-Image, a free tool that runs on your home PC and finally fixes the biggest AI art problem: messy text.

Founders, Coaches & Business Owners - Listen up!

You're the Bottleneck in Your Content System

Most founders reading this know content matters. But between calls, product work, and actually running the business where's the time to record, edit, and post?

Here's the truth: your personal brand can grow without you being in front of a camera daily.

I run the Founder's Clone Challenge an 8-day system that teaches you to build an AI-powered content workflow. By Day 8, you have a digital clone (your voice, your face, your style) that your editor can use to create daily content while you focus on what actually moves your business forward.

100+ founders have already built theirs. Batch 5 starts on December 5.

→ Learn more at foundersclone.com