- unwind ai
- Posts
- China's New 1 Trillion Model Qwen3 Max
China's New 1 Trillion Model Qwen3 Max
PLUS: Build your own Claude Code agent, Two stealth models with 2M context
Shubham Saboo & Gargi Gupta
September 08, 2025
Today’s top AI Highlights:
& so much more!
Read time: 3 mins
AI Tutorial
We have created a complete Google Agent Development Kit crash course with 9 comprehensive tutorials!
This tutorial series takes you from zero to hero in building AI agents with Google's Agent Development Kit.
What's covered:
Starter Agent - Your first ADK agent with basic workflow
Model Agnostic - OpenAI and Anthropic integration patterns
Structured Output - Type-safe responses with Pydantic schemas
Tool Integration - Built-in tools, custom functions, LangChain, CrewAI, MCP
Memory Systems - Session management with in-memory and SQLite storage
Callbacks & Monitoring - Agent lifecycle, LLM interactions, tool execution tracking
Plugins - Cross-cutting concerns and global callback management
Multi-Agent Patterns - Sequential, loop, and parallel agent orchestration
Each tutorial includes explanations, working code examples, and step-by-step instructions.
Everything is 100% open-source.
We share hands-on tutorials like this every week, designed to help you stay ahead in the world of AI. If you're serious about leveling up your AI skills and staying ahead of the curve, subscribe now and be the first to access our latest tutorials.
Latest Developments
NVIDIA just released a Deep Research system where you write the strategy, and it writes the code to execute it.
Most research agents follow one hardcoded approach - search, analyze, report. NVIDIA's Universal Deep Research (UDR) lets you write the research process itself in plain English, then automatically converts your methodology into executable Python code that runs in a sandbox. You literally tell the system "first do this, then do that, then compile results this way," and UDR generates working code that follows your exact instructions step-by-step.
You can define how many search rounds to do, what validation steps to include, and how to structure the final report. The system wraps around any language model, so you can pair the latest models with your custom research strategies without being tied to specific providers.
Key Highlights:
Code generation from strategy - Write your research methodology in plain English and UDR automatically converts it into executable Python code that implements your exact approach.
Sandboxed execution - Your custom-generated code runs in an isolated environment where it safely performs searches, processes data, and builds reports according to your specifications.
Complete process control - Define every aspect from search patterns and validation steps to report structure, rather than accepting predetermined workflows.
Real-time monitoring - Watch your generated code execute with live progress updates showing each step of your custom research process.
Open-source code - Both the FastAPI backend and Next.js frontend are available on GitHub, so you can deploy your own instance or contribute to the framework development.
Building your own CLI coding agent sounds like reinventing the wheel when Claude Code, OpenAI Codex CLI, Gemini CLI, etc. already exist.
But this developer from Martin Fowler’s team discovered something important: the magic isn't in any single AI model, but in how you architect the toolchain around it. His step-by-step build reveals the specific design patterns that make commercial agents so effective.
The article walks through building a CLI agent using Pydantic-AI and the MCP, starting with simple test execution and evolving into a system that can debug, reason, and make surgical code edits across entire codebases. What makes this particularly valuable is his focus on the incremental capabilities that turn an AI assistant from just helpful to genuinely collaborative. The agent they built doesn't just generate code; it runs tests, searches current documentation, accesses AWS services, and maintains context across complex multi-step workflows.
Key Highlights:
MCP servers are almost everything - MCP tools allows agents to access specialized tools through a standardized interface, making capabilities like sandboxed Python execution, up-to-date documentation, and AWS services easily pluggable without custom integrations.
Context is king - CLI agents are great partly because they maintain state across test runs, file edits, and documentation searches, creating a genuinely collaborative debugging experience.
Specialization beats generalization - His custom agent outperformed general-purpose tools on their specific codebase because it understood their development patterns, test-driven approach, and tool preferences rather than trying to be everything to everyone.
Structured reasoning is crucial - Adding systematic problem-solving capabilities transformed the agent from generating quick fixes to breaking down complex issues, exploring alternatives, and explaining its thought process like a senior developer.
Follow this implementation guide and build your own agent over the coming weekend. This will give you insights into agent architecture that might change how you use commercial CLI tools.
It’s go-time for holiday campaigns
Roku Ads Manager makes it easy to extend your Q4 campaign to performance CTV.
You can:
Easily launch self-serve CTV ads
Repurpose your social content for TV
Drive purchases directly on-screen with shoppable ads
A/B test to discover your most effective offers
The holidays only come once a year. Get started now with a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.
Quick Bites
Another 1 Trillion Model from China after Kimi K2 (but not open-source)
Alibaba Qwen team teased their new flagship model Qwen3 Max, their biggest model yet, with 1T parameters and a 262K context window. Since it’s from the Max series, the model is neither open-source nor open-weight - it’s available on Qwen Chat and via API only. The model shows substantial improvements over the Qwen 2.5 series across Chinese-English comprehension, complex instruction following, and agentic tasks, while reducing hallucinations. Other details on the model are under wraps.
Two Stealth Models on Open Router with a 2M Context Window
OpenRouter just dropped two "stealth" general-purpose models called Sonoma Alpha - Dusk (fast) and Sky (maximum intelligence), both packing a massive 2 million token context window, supporting image inputs, parallel tool calling, and currently free during alpha testing. With no disclosed origins, the AI community is buzzing with theories - it’s either xAI (potentially the next Grok iteration) or Google (possibly Gemini 3 Pro and Flash), based on performance patterns and stylistic fingerprints in outputs. The early feedback from users is largely positive, especially in reasoning and coding tasks. Worth a spin while they're still free and anonymous!
The Most Hands-on Guide to Agentic System - 100% Free
This Distinguished Engineer at Google has released a free 424-page book, Agentic Design Patterns. The book covers 21 agentic design patterns, spanning everything from prompt chaining, tool use, and multi-agent collaboration to self-correction. The book also covers practical code examples across LangChain, LangGraph, Crew AI, and Google's Agent Developer Kit, with each pattern getting its own dedicated chapter. The ebook is completely free, while a paperback version is available on Amazon.
GLM 4.5 $3 Subscription Might Make You Ditch $200 Claude Max
China’s Z.ai has released new dedicated coding plans to use their GLM-4.5 model exclusively with AI coding agents like Claude Code, Roo Code, Cline, etc. The Lite subscription is for$3/month and the Pro for $15/month. With 120-600 prompts per 5-hour cycle respectively, it's capturing developers who find Claude's $200 Max plan excessive for coding needs. The timing is very smart - GLM-4.5 is one of the best coding models globally and offers significantly more usage than Claude's tiers. Some developers have already downgraded their Claude subscriptions!
Tools of the Trade
System Prompts and Models of AI Tools - This open-source repository has entire system prompts, tools, and AI models used for the most popular agentic tools, including Claude Code, v0, Manus AI, Devin, Replit, Warp, Trae, etc.
DeepMCPAgent - Build AI agents that auto-discover and use tools via MCP. No manual wiring, just connect your LangChain model (OpenAI, Anthropic, Ollama, etc.) and the agent dynamically generates typed tools for production-ready workflows.
Codex-subagents-mcp - This MCP server adds Claude Code-style sub-agents to OpenAI's Codex CLI, which lacks this feature natively. It works by loading agent personas from agents.md/.json files and executing tasks through isolated temporary workspaces using Codex profiles.
Awesome LLM Apps: A curated collection of LLM apps with RAG, AI Agents, multi-agent teams, MCP, voice agents, and more. The apps use models from OpenAI, Anthropic, Google, and open-source models like DeepSeek, Qwen, and Llama that you can run locally on your computer.
(Now accepting GitHub sponsorships)
Hot Takes
I think congrats again to OpenAI for cooking with GPT-5 Pro. This is the third time I've struggled on something complex/gnarly for an hour on and off with CC, then 5 Pro goes off for 10 minutes and comes back with code that works out of the box. I had CC read the 5 Pro version and it wrote up 2 paragraphs admiring it (very wholesome). If you're not giving it your hardest problems you're probably missing out. ~
Andrej KarpathySoftware engineering is 90% database and 10% frontend. ~
Armin Ronacher
That’s all for today! See you tomorrow with more such AI-filled content.
Don’t forget to share this newsletter on your social channels and tag Unwind AI to support us!
PS: We curate this AI newsletter every day for FREE, your support is what keeps us going. If you find value in what you read, share it with at least one, two (or 20) of your friends 😉
Reply