Good Morning! Here is today's breakdown:
Google Gemma 4 now runs free on your laptop with no internet needed
Ideogram 4.0 just became the best open-source image model in the world
Reve 2.0 launched the best 4K image model anyone can use today
The first mistake most people make with Claude
4 new AI tools worth trying today
AI AGENTS
Google released Gemma 4 12B, a free open-source multimodal model that runs entirely on a standard 16GB laptop with no cloud connection required, handling text, images, video, and audio together in one unified architecture for the first time in a model this size. Available now on Hugging Face under Apache 2.0, commercial use permitted.
Runs locally on any device with 16GB of VRAM or unified memory with no internet after download, and as low as 8GB in quantized form, with same-day support across Ollama, vLLM, llama.cpp, and MLX
Encoder-free architecture feeds all four modalities directly into one LLM backbone with no separate processing pipelines, reducing both latency and memory use compared to traditional multimodal designs
First medium-sized Gemma model with native audio input, plus a 256K context window, thinking mode via the <think> token, and native function calling for agentic workflows
Apache 2.0 license means free to download, run locally, and use commercially with no restrictions beyond attribution
Local AI has been split into two camps: small models that run on personal hardware but miss half of what frontier AI can do, and frontier models that handle everything but live in the cloud. Gemma 4 12B closes that gap. It is the first open model at this size that handles all four modalities together on hardware most people already own. The Apache 2.0 license removes the commercial restriction that made previous local models difficult for production use. For anyone building privacy-sensitive applications, running AI in secure offline environments, or wanting to stop routing sensitive documents through cloud APIs, this is the most practically significant open model release of 2026 so far.
AI AGENTS
Ideogram has launched Ideogram 4.0, the company's first ever open-weight text-to-image model, with weights free to download from GitHub and Hugging Face for anyone to run on their own hardware and fine-tune on their own data. The 9.3B parameter model ranks #1 among all open-weight models on DesignArena and is live on every Ideogram plan and the API today.
Weights are on GitHub at ideogram-oss/ideogram4 and Hugging Face in fp8 and nf4 versions; personal and research use is free, commercial use requires a paid license, and fine-tuning on your own brand data is fully supported
Native 2K resolution output with transparent background support, precise layout control via bounding boxes, and 47.9% text rendering accuracy rated by professional designers, making it production-ready for logos, posters, and brand assets
Trained from scratch on a structured JSON prompting interface using a single-stream Diffusion Transformer with Qwen3-VL-8B-Instruct as the text encoder, not a fine-tune of any existing model
Ranks #1 on DesignArena among all open-weight models and #8 overall, behind only closed models from OpenAI and Google, with editable text and movable layers coming soon
Ideogram built its reputation on one thing: accurate text inside images, the capability every other image model gets consistently wrong. 4.0 adds bounding-box layout control, native 2K output, and transparent backgrounds, turning it from an image generator into a production-grade design tool. The open-weight release is the bigger signal. Ideogram has been one of the most design-centric closed platforms in the space. Releasing the weights changes who can build with it and how. For any team running brand tooling, design automation, or marketing production that requires readable text in generated images, this is now the reference open model.
AI RESEARCH
Reve has launched Reve 2.0, a 4K image model that replaces text prompts with layout representations where every image element carries a location, size, and description, making it the first image model where composition is directly editable rather than re-generated from scratch. It ranks #2 on the global Image Arena leaderboard, behind only OpenAI's GPT Image 2.
Layout representations replace traditional text prompts: each image element is defined by position, size, and description in a structured hierarchy that users refine through natural language or by directly editing the layout, with CLIP similarity rising from 0.865 at zero regions to 0.929 at 50 regions
Ranks #2 on Image Arena with a score of 1,280 from 3,455 votes, behind only OpenAI GPT Image 2 and ahead of Google Gemini 3.1 Flash Image Preview, despite being trained on 10x fewer GPUs than comparable frontier models
Native 4K output with no separate upscaling step, making generated images production-ready for ads, pitch decks, landing pages, and print assets directly out of the model
Available now at reve.art with precise composition control via the Reve editor, designed for both human designers and agentic workflows that need to modify specific image regions without regenerating the whole image
Every text-to-image model has the same core problem: you describe what you want, the model makes a guess, and you start over when it misses. Reve 2.0 addresses that at the architecture level by making the internal image representation a structured, editable object. If composition is a data structure rather than a prompt, both humans and AI agents can modify specific elements without touching the rest. That changes image generation from a one-shot output into something closer to an editable creative file, which is how production design work actually needs to behave. The #2 Arena ranking behind only OpenAI validates the approach.
HOW TO AI
📋 How to generate production-ready images with accurate text using Ideogram 4.0
Ideogram has been the only AI image tool that actually gets text right inside images. Logos, posters, packaging, social ads, everything every other model gets wrong, Ideogram gets right. Version 4.0 just added native 2K output, precise layout control via bounding boxes, and transparent backgrounds. Here is how to use it from scratch.
STEP 1: Get access
Head to ideogram.ai and create a free account. Ideogram 4.0 is live on every plan including free today. If you want to download the weights and run locally, the model is on GitHub at ideogram-oss/ideogram4 and on Hugging Face.
STEP 2: Write your first prompt
Start with plain natural language:
"A minimalist poster for a coffee brand called Morning Ritual. Clean sans-serif typography. Warm earth tones. The text Morning Ritual centered at the top."
The model renders the text accurately inside the image. This is the capability that has broken every other image model for years.
STEP 3: Use layout control to place elements precisely
Switch to Layout mode in the interface. You can now:
Draw a box where you want the headline to appear
Draw a separate box for a logo or product image
Set color palette constraints per region
Define what goes in each region with a short descriptionSTEP 4: Turn On Meeting Bots
The model follows the layout instructions precisely instead of guessing from a text prompt. This is what makes 4.0 practical for real brand work.
STEP 4: Export in 2K for production use
Ideogram 4.0 generates natively at 2K resolution. No upscaling required. Download directly from the interface ready for ads, pitch decks, print materials, and landing pages. Toggle transparent background before generating to get a PNG with no background, ready to drop into any design tool.
PREMIUM GUIDE
📋 The first mistake most people make with Claude
Most people use Claude like a smarter Google search.
Open a chat. Type a question. Get an answer. Close it.
That approach works for one-off tasks. It fails completely for anything you do repeatedly.
The first mistake most people make, and it's everywhere.
Before anything else: go to Settings and look at the fields that say "tell Claude about yourself" and "personal preferences Claude should consider."
Everyone recommends filling these out. Don't.
Here's why:
These instructions apply to every single conversation, across every topic, no matter the context. If you use Claude across a wide range of tasks, writing, research, analysis, coding, a single instruction set doesn't fit all of them. You'll end up with weird, off-tone responses in the wrong situations.
Leave these completely blank.
Set up context in projects instead. Projects let you give Claude specific instructions scoped to a specific type of work. The right instructions apply at the right time. It's a dramatically better system.
Projects, the most impactful change you can make.
Projects are self-contained workspaces with their own memory, chat history, knowledge base, and custom instructions.
Each project is a dedicated environment for a specific type of work. The custom instructions you set only apply inside that project. The files you upload are always available in that project. Every conversation inside inherits all of that context automatically.
No re-explaining. No re-uploading. No re-writing context from scratch.
How to set one up:
Click New Project. Give it a name.
Go to Instructions and write what Claude should know and how it should behave inside this project.
Go to Project Files and upload any documents Claude should always have access to.
What goes in the instructions:
What this project is about
How Claude should approach tasks here
Tone, style, and format preferences
Any specific rules or constraints
What goes in the files:
Brand guidelines
Style guides
Examples of past work you want Claude to emulate
SOPs, templates, reference documents
One project for writing. One for research. One for client work. One for your own content. Each one gets the right context for the right job.
Here is what a real instruction set looks like for a newsletter project:
"You are my newsletter assistant. Help me with any newsletter-related task: ideation, angle selection, drafting, editing, and repurposing. You have access to my folder with past newsletters for tone reference, my ICP doc, my voice personality guide, and my newsletter strategy doc. Always mimic my tone from the examples. Never sound generic. Always read the reference files before executing."
Want to read the full guide?
Become a paying subscriber to get access to all premium content.
✓ Full archive of premium guides with ready-to-use prompts
✓ Structured AI courses (step-by-step, start-to-finish)
✓ Every upcoming premium tutorial
Miso Labs launched Miso One on June 3, an 8B open-source text-to-speech model with 110ms latency and real emotional range: warmth, hesitation, excitement, grief. Model weights are free on GitHub at MisoLabsAI/MisoTTS.
Google Labs released Dreambeans on June 3, an experimental mobile app that connects to all your Google apps using Personal Intelligence to deliver personalized story collections every day, surfacing things you would otherwise miss.
Nitrosend launched on June 3 as the first email platform with no dashboard. Manage your entire email operation from Claude, ChatGPT, Codex, or Gemini via MCP. One prompt sends to 10,000 people.
🤖 Google Gemma 4 12B: Google's free open-source model that handles text, images, video and audio on your laptop with no internet. Apache 2.0. Download and run today.
🎨 Ideogram 4.0: The best open-source image model in the world. Native 2K resolution, precise layout control, accurate text in images. Free to use today.
🖼️ Reve 2.0: The best 4K image model in the world. Layout-based generation that makes every image element editable. Live now.
🎙️ Miso One: Open-source TTS model with real emotional range and 110ms latency. Free weights on GitHub. More human than anything else available.
THAT’S IT FOR TODAY
Thanks for making it to the end! I put my heart into every email I send, I hope you are enjoying it. Let me know your thoughts so I can make the next one even better!
See you tomorrow :)
- Dr. Alvaro Cintas