Hey there, Nomadev here! If you're reading this, you're probably as excited about AI as I am. Every week brings something new, with fresh models, smarter agents, and ideas that push the limits of what we can build.
So I thought I’d make it easier for you to stay in the loop.
Here’s a quick roundup of everything that happened in AI this week, covering updates from OpenAI, Anthropic, Nvidia, IBM, and more.
Whether you’re a developer, a researcher, or just an AI enthusiast who loves seeing how fast this space moves, this is your quick catch-up. Let’s dive in! 🚀
1. OpenAI Sora 2
Sora 2 is OpenAI’s next big leap in video generation, turning simple text prompts into high-fidelity, cinematic visuals.
Whether you want to create storyboards, concept videos, or just explore visual storytelling with AI, Sora 2 makes it feel effortless.
This isn't just an upgrade, it's a preview of how creative workflows will change.
Here’s how Sora 2 makes things magical:
- Realistic Motion: Smooth, consistent movement across scenes with fewer artifacts.
- Scene Coherence: Objects and subjects stay consistent throughout the generated video.
- Camera Control: Prompts can influence angle, zoom, and transitions for more directed storytelling.
- Better Prompts = Better Results: Try describing mood, lighting, and movement for high-quality outputs.
- Built for Creators: Ideal for product ads, mood boards, music videos, and explainer visuals.
OpenAI is gradually rolling it out to trusted partners.
In the meantime, follow the latest updates from OpenAI on X and start crafting prompts of your own.
Sora 2 isn't just a video tool, it's your creative assistant on turbo mode.
2. Qwen3-VL-30B-A3B Instruct & Thinking
Alibaba’s Qwen team launched Qwen3-VL-30B-A3B Instruct & Thinking, a new generation of large vision-language models.
With just 3B active parameters, it delivers powerhouse performance while staying lightweight and efficient.
This release brings all the capabilities of Qwen3-VL in a smaller but sharper package.
Here’s why Qwen3-VL-30B-A3B stands out:
- Big performance, small footprint: 3B active parameters rival GPT‑5 Mini and Claude 4 Sonnet on many benchmarks.
- Wide skill range: STEM problem solving, VQA, OCR, Video understanding, and Agent tasks.
- Consistent wins: Outperforms bigger models across multiple evaluation benchmarks.
- All-in-one model: Instruct and “Thinking” variants available for different reasoning tasks.
- Optimized for developers: A more accessible, high-quality open model for experimentation and deployment.
Qwen3-VL-30B-A3B isn't just another model release, it's a serious competitor bringing high-end performance into a smaller footprint.
Follow their updates on Qwen's official X account to stay ahead.
3. AntLing Ring-1T: 1 Trillion Open-Source Thinking Model
AntLingAGI just launched Ring-1T, the first open-source model with 1 trillion parameters focused on deep thinking and reasoning tasks.
Early benchmarks show strong results across math, logic, and natural language problems, with impressive performance on high-difficulty benchmarks.
Here’s why Ring-1T is a big deal:
- 1 Trillion Parameters: One of the largest open-source models ever released.
- Strong benchmarks: Scored 92.6 on AIME25, 84.5 on HMMT25, and 94.7 on CF (Competition Format) tasks.
- IMO-level math: Solved IMO25 Q3 in one shot, and showed progress on Q1, Q2, Q4, and Q5.
- Open and evolving: Still improving, and already competitive with top models like GPT-5 Thinking and DeepSeek Terminus.
- Designed for thinkers: Built for reasoning-heavy tasks like coding competitions, math olympiads, and logic-based prompts.
If you're exploring reasoning agents or want to go beyond general chat models, Ring-1T is worth a test run.
4. Claude Sonnet 4.5: Best Model for Agentic Coding
Anthropic launched Claude Sonnet 4.5, their strongest model yet for building agentic systems and reasoning with computers.
It’s designed specifically to help developers create smarter agents, solve logic-heavy tasks, and write more reliable code.
Here’s what makes Claude Sonnet 4.5 stand out:
- Top-tier agentic coding: Scored 77.2% on SWE-Bench (code reasoning) and 82.0% when allowed to use external tools.
- Big jump in logic tasks: Terminal-Bench, high school math, and graduate-level reasoning scores have all improved.
- Tool-use ready: Claude Sonnet 4.5 shows strong performance in l2-bench, which measures how well agents use tools.
- Best at math and reasoning: Scored 100% on AIME 2025, 83.4% on GPAQ, and leads in multilingual Q&A tasks.
- Built for real-world agents: Especially good at using computers, coding agents, and handling complex toolchains.
If you're building serious AI agents or systems that interact with files, tools, or structured data, Claude Sonnet 4.5 is worth a deep dive.
Check out Anthropic's official post for more benchmarks.
5. GLM-4.6: Agentic, Reasoning, and Coding Powerhouse
Zhipu AI introduced GLM‑4.6, a flagship model built for real-world coding, reasoning, and long-context understanding.
It’s designed for developers building advanced agentic systems, and performs especially well in tasks that require tool use, search, and logic.
Why GLM‑4.6 is worth a look:
- Agent-ready: Optimized for agentic workflows, coding, and task planning.
- Handles long context: Supports input lengths of up to 200,000 tokens.
- Competitive benchmarks: Scores on AIME 25, GPQA, HLE, and Terminal-Bench match or exceed Claude Sonnet 4.5 and DeepSeek.
- Strong in reasoning and tools: Excels in long reasoning chains, tool use tasks, and multistep problem solving.
- Built for devs: Targeted toward real coding and automation use cases, not just chatbot-style interaction.
You can follow updates and access the API from Zhipu's official post.
GLM‑4.6 is shaping up to be one of the most versatile open models in the agent ecosystem.
6. Coral v1: Launch and Monetize AI Agents
Coral Protocol launched v1, unlocking a full-stack system to orchestrate, monetize, and deploy AI agents at scale.
It aims to power the “agent economy” by letting anyone turn AI agents into reusable and monetizable software units.
Here’s how Coral v1 makes it happen:
- No lock-in: Reuse AI agents built with any framework or language.
- Launch-ready: Deploy agents without having to build infrastructure from scratch.
- Open marketplace: Publish agents for others to use, collaborate, or buy.
- Built-in monetization: Sell or license your agents and earn from contributions.
- Built for scale: Perfect for solo devs and startups looking to ship quickly.
If you're working on AI products and want to plug into a growing ecosystem, Coral v1 might be the missing link.
7. NotebookLM: Customizable Chat Experience
NotebookLM introduced a new personalization layer for its chat interface, letting users tailor the conversation flow to better fit their learning and working styles.
Whether you're using it as a study guide or an assistant, it now adapts more deeply to your preferences.
Here’s what’s new in NotebookLM:
- Custom response length: Choose how short or long the answers should be.
- Adjustable conversation style: Toggle between concise, explanatory, or question-driven replies.
- Learning Guide mode: A new style that tests and reinforces your understanding of the material.
- Focus on depth: Designed to go beyond summarization and promote deeper reasoning.
- Great for students and researchers: Ideal for learning workflows or guided comprehension.
You can explore the updates from NotebookLM's announcement here.
This update brings NotebookLM closer to becoming a personalized tutor in your pocket.
8. Comet by Perplexity: Now Available Globally
Comet, Perplexity’s personal AI assistant, is now officially available to everyone worldwide.
After spending 84 days in waitlist mode, millions of users have joined to explore a new way of using the internet.
Here’s what Comet brings to the table:
- Personal AI assistant: Designed to help you search, think, and learn faster online.
- Conversational search: Combines real-time information retrieval with chat-style interaction.
- Waitlist is over: No more invites needed. Anyone can try it out right now.
- Focus on usability: Built for everyday internet users who want smarter answers, faster.
- AI-enhanced browsing: Think of it as a more intelligent search engine that talks back.
You can watch the official announcement and explore Comet from Perplexity's post.
Whether you're researching or just curious, Comet gives you a fresh way to browse.
9. DeepSeek-V3.2-Exp: Faster, Smarter, Cheaper
DeepSeek released V3.2-Exp, their latest experimental model built on the V3.1-Terminus architecture.
It introduces DeepSeek Sparse Attention (DSA), enabling faster inference and training, especially for long-context processing.
Here’s what’s new:
- Sparse Attention mechanism: Improves speed and efficiency on large inputs.
- Live across platforms: Available now via App, Web, and API.
- 50%+ cheaper: Major price cuts for API usage.
- Focused on real-world use: Great for apps needing fast and scalable model calls.
If you're optimizing cost, speed, and context length in your LLM workflows, DeepSeek-V3.2-Exp is worth exploring.
10. Granite 4.0 by IBM: Lightweight Models for Local Use
IBM just released Granite 4.0, a new series of small language models built for agentic tasks, RAG, and document analysis.
What’s most exciting is the Micro (3.4B) variant that runs entirely in your browser using WebGPU, with no server needed.
Why Granite 4.0 matters:
-
Runs locally: The Micro model can run 100% on-device in your browser with
Transformers.js
. - Privacy-friendly: No data is sent to a server, and it can work offline after loading.
- Fast and efficient: Designed for quick in-browser inference and lightweight deployments.
- Great for edge cases: Perfect for building low-latency, privacy-sensitive apps.
If you're looking for compact models that work offline, Granite 4.0 opens up exciting new possibilities.
11. Nano Banana by Google DeepMind: Production-Ready Image Model
Google DeepMind released a full guide for using Nano Banana, a production-ready image generation model that’s part of the Gemini 2.5 Flash stack.
It’s designed for developers who need dynamic image outputs with more control and flexibility.
Why Nano Banana stands out:
- Image-only mode: Generate just visuals without extra text or padding.
- Creative freedom: Specify aspect ratios and fine-tune composition.
- Built for scale: Suitable for production environments, not just demos.
- Gemini-compatible: Works seamlessly within Google’s Gemini AI ecosystem.
If you're building visually dynamic apps or want to embed reliable image generation into your product flow, Nano Banana is worth checking out.
12. Unlock the Full Potential of Your Mac with Spec
Spec brings AI memory and smart automation to your Mac.
It proactively drafts replies, summarizes documents, and helps you stay organized across all your tools — before you even ask.
Why Spec feels like a brain for your Mac:
- Cross-app memory: Connects knowledge across iMessage, Slack, email, and calendar.
- No more context switching: Everything feels unified and accessible.
- Helpful before you ask: Anticipates needs and assists proactively.
- Boosts productivity: Great for people juggling comms, docs, and deadlines.
If your daily workflow lives across multiple apps, Spec might be the assistant you didn't know you needed.
13. Composer: The First AI Agent for Document Processing
Composer is built to tackle one of the most common but painful workflows — document processing.
It promises production-grade accuracy in under 10 minutes with no complex setup.
Here’s why Composer is exciting:
- 99% accuracy: Some early users reported near-perfect results on schema-heavy docs.
- Minutes, not hours: Gets up and running in under 10 minutes.
- Built for scale: Designed for teams that process lots of structured documents.
- Agentic optimization: Learns and adapts to recurring document types and formats.
If you're in finance, HR, legal, or ops — Composer could be your new best friend.
14. C1 by Thesys: Generative UI for LLMs
C1 is a powerful new API that lets LLMs respond with rich interactive UIs, not just plain text.
Created by Thesys, it aims to change how AI interfaces work — moving beyond chat into charts, forms, and cards.
What makes C1 different:
- Interactive responses: LLMs can now return UI components like charts or inputs.
- Great for apps: Useful in dashboards, admin panels, education tools, and assistants.
- Plug-and-play: Easily integrates with existing LLM pipelines.
- Launch perks: They offered up to 5M tokens free on Product Hunt.
If you've been dreaming of "ChatGPT meets Notion-style UI," this API brings it closer to reality.
15. CrewAI Launches AMP: The OS for AI Agents
CrewAI just launched AMP, their Agent Management Platform, designed to be the operating system for AI agents in production.
It’s already seeing major adoption with Fortune 500 companies and public use cases.
What AMP brings to the table:
- Agent OS: Centralized interface to deploy, monitor, and manage AI agents.
- Massive scale: 100K+ executions in 15 days, 30+ live use cases.
- Enterprise-ready: Used by large corporations and public companies.
- Workflow visualizer: Clean UI to build complex agent pipelines visually.
If you’re scaling agentic systems and want production-level observability and orchestration, AMP is built for you.
15. CrewAI Launches AMP: The OS for AI Agents
CrewAI just launched AMP, their Agent Management Platform, designed to be the operating system for AI agents in production.
It’s already seeing major adoption with Fortune 500 companies and public use cases.
What AMP brings to the table:
- Agent OS: Centralized interface to deploy, monitor, and manage AI agents.
- Massive scale: 100K+ executions in 15 days, 30+ live use cases.
- Enterprise-ready: Used by large corporations and public companies.
- Workflow visualizer: Clean UI to build complex agent pipelines visually.
If you’re scaling agentic systems and want production-level observability and orchestration, AMP is built for you.
That’s it for this week’s AI updates.
If you found something useful in here, feel free to drop a message or tag me on X. I’ll be doing this every week, so you can always come back for a quick catch-up.
Follow @thenomadevel on X for the full thread and more updates like this.
See you next week.
Top comments (0)