Edition #7. The week of NVIDIA GTC. Three developments that together show where AI agents now stand.
MiMo-V2: the first model built for the Agent era
One of the engineers behind DeepSeek R1 launched MiMo-V2. Not just an update, but a complete model family specifically designed for AI agents. Pro, Omni and TTS versions. Built on a 1-trillion-parameter base model with a 1 million token context window.
This is exactly what's been observed in recent months. The models from a year ago were good at having conversations. The models of today are good at executing tasks. That's a fundamental difference. ChatGPT was a conversation partner. MiMo-V2 is a coworker.
The model is optimized for complex agentic workflows where an agent combines multiple tools, sources and steps to complete a task. Ultra-low latency, 1M context window, and it goes open source as soon as it's stable enough.
The data confirms this. On OpenRouter, the platform where developers compare and use AI models, Hunter Alpha, the model behind MiMo-V2, sits at number 1. 1.51 trillion tokens processed. A growth of 2,017 percent this week.
What I did with AI agents this week
This week, Atlas, my own agent, processed the full administration of two BVs and researched dozens of potential customers. That's an AI agent. Not a chatbot. A digital employee that executes tasks.
The real power sits in the combination with custom software. Fully configured agents on dedicated servers. Not a bare install, but working digital employees with their own tools. Think of a sales agent with its own CRM that finds leads for you daily, analyses websites, and writes personalized emails. Or a bookkeeping agent that's available 24/7, proactively processes your receipts, calculates VAT, and keeps your administration up to date before you even think of it.
No sick days. No vacations. No salary negotiations. Always available, always consistent, and a fraction of the cost of an employee.
