noticias radar ia modelos ia pesos abiertos automatizacion

Yesterday's radar was about putting a leash on agents; this weekend the question is different: who owns the AI your company uses? This week Alibaba is going to release as open weights a model that competes with the best from OpenAI and Anthropic —downloadable for you to host yourself—, Meta is starting a price war with its first programming agent, and even OpenAI has had to freeze its most powerful model for fear of what it can do. In parallel, the first major ChatGPT usage study shows where the real money is: not in science fiction agents, but in automating the writing and paperwork your team already does by hand. This weekend's takeaway: you have more and more room to choose —and even own— your AI by model, price, or control, but the safe return remains in boring, repeatable tasks.

Soon you'll be able to run world-class AI on your own servers, without a token bill

On August 3, Alibaba presented Qwen3.8-Max, its most capable model to date —a giant with 2.4 trillion parameters, a one-million-token window, and performance that the company claims rivals the best from OpenAI and Anthropic— and confirmed that this week it will publish its open weights, along with a 27-billion-parameter sibling, on Hugging Face and ModelScope. This is the first time it has released a high-end model. For your company: open weights mean that you or your provider can host the model on your own infrastructure —sensitive data stays at home (key for the AI Act and AESIA) and the token bill disappears—; however, the 2.4 trillion giant is a data center job, so the one you can really self-host is the 27 billion one, and it’s wise to wait and see what license it launches with before building anything on top of it. Source

Coding with AI gets cheaper: Meta starts a price war with its first code agent

On August 5, Meta launched Muse Code, its first programming agent: a command-line tool that plans, writes, and validates code from start to finish and, for large projects, distributes the work among several agents running simultaneously in isolated environments. Its hook is the price —below what Anthropic or OpenAI charge— with two modes: pay-per-use or a cheaper "contributor" rate in exchange for letting Meta train on your code. For your company: if your business develops or maintains software, assisted programming is no longer just autocomplete, but delegating entire tasks, and the competition is driving prices down in your favor; but watch the fine print, because the cheap rate is usually paid for with your data: if your code is sensitive, demand zero data retention —which Meta already offers to enterprises— before plugging anything in. Source

OpenAI shut down its most powerful model: the newest isn't always the one you should use

On August 7, OpenAI revealed that its upcoming model, Astra, could reach a "critical" level of offensive cybersecurity capability —finding and independently exploiting zero-day vulnerabilities in real systems, and even orchestrating full cyberattacks without human intervention— the first time its own safety framework has triggered the maximum brake. It has limited work with the model, locked it in isolated environments with total surveillance, and will not deploy it until controls are strengthened. For your company: more capability doesn't mean more ready to use; choose the model based on what your task needs —often a smaller, cheaper one does it just as well and with less risk— not because it's the newest or biggest. And, as we saw yesterday, any powerful automation goes in a box: minimum permissions, isolated, and with human approval for whatever goes out to the world. Source

What really pays off: automating the writing and paperwork you already do by hand

On August 6, OpenAI published its first large-scale study of ChatGPT usage, with clear data: at work, people use it to do or create something —not just to ask— more than twice as much as outside of work; the star use is writing and, within that, mostly editing, reviewing, and translating existing texts rather than drafting from scratch; and the general pattern is moving from "asking" to "doing". For your company: the biggest return isn't in the flashiest agent, but in those boring and repeatable text tasks your team already does every day —drafting responses, reviewing, summarizing, translating—; choose one, turn it into a template or a fixed workflow this week, and measure the time it saves you before thinking about anything more ambitious. Source

What to watch this week?

First, if Alibaba delivers and releases the Qwen weights —and under what license— because it marks how far you can bring AI into your own home; second, more "own your model" moves and price wars in agents, and how regulators and labs respond to OpenAI freezing Astra. And a question for Monday: of everything your team writes by hand each week, which is the first task you are going to turn into a template?

Watch the 1-minute video