Alibaba broke the last lock left on its catalog: Qwen3.8-Max, its largest model, will be the first in the Max line to publish its weights, according to details published by the South China Morning Post. “Open weights” means anyone can download the model and run it on their own hardware, without depending on anyone else’s server. It is a model with 2.4 trillion total parameters (about 95 billion active per query) and a one-million-token window, designed for tasks that last days: software development, research reproduction, chip design. The weights arrive in mid-August.
For Latin America, the number that matters is not the parameter count but the jurisdiction. For the first time, a model with frontier capability can run on one’s own infrastructure, with the data inside the country and without paying a closed lab’s rates, which is exactly the argument that currently holds up much of the public procurement of AI in the region. In the meantime it is already available via API at $2 and $6 per million input and output tokens (at 10% of that rate during the preview), and it speaks OpenAI’s and Anthropic’s API formats, so migrating costs no more than changing an address. A caveat is in order: the comparisons that place it close to or above Claude Opus 4.8, Claude Fable 5 and GPT-5.6 Sol come from the company’s own internal tests, without independent verification.
The same day, Rest of World published a map of the political fight all this is setting off: 179 Silicon Valley startups asking Washington to preserve access to Chinese open models, and Anthropic pushing in the opposite direction. The region is the market that this debate takes for granted, and it has no one sitting at the table: if the United States restricts, it restricts precisely the input that today makes AI viable on Latin American budgets.
Also today
- AWS spells out what it looks at before installing a data center in Latin America — Regulation, taxes, political stability and demand. None of the four criteria belongs to the ministry where the region usually houses its AI policy. The company has $4 billion under construction in Chile, a local zone in Bogotá for this year and projected zones in Buenos Aires and Rio de Janeiro.
- INTERPOL links AI to more than half of reported cybercrime in Africa — 55% of cases across 36 countries, losses that went from $192 million to $484 million in one year, and 600,000 extortion cases using deepfakes and synthetic identities that defeat biometric verification. There is no equivalent report for our region.
- Two teams solve the same open problem in quantum cryptography and publish three hours apart — Both used GPT-5.6. When everyone works with the same tool, simultaneous discovery stops being a historical anecdote and becomes the normal mode of production; they are considering merging the two papers.
- The U.S. Congress spent $113,740 on AI in one year, and 90% went to ChatGPT — 798 ChatGPT transactions against 37 for Claude, with Democratic offices spending more than three times as much as Republican ones. It is the first public record of a legislature’s actual AI spending, and it is microscopic.
- IBM and the Ponemon Institute publish the record cost of a data breach — A global average of $4.99 million, 12% more than the year before. The actionable figure: 92% of companies that suffered breaches in AI systems lacked adequate access controls.
In the region
The week begins with two regional moves that do not come from where people usually look for them. Panama presented its National Advanced Technologies Agenda, which brings together in a single instrument the National AI Strategy and the National Semiconductor and Microelectronics Strategy, with technical coordination by SENACYT, six pillars, about $105 million for the hardware component and a newly created National Commission on Critical and Emerging Technologies. It is the first country in the region to treat software and hardware as a single policy, and it deliberately chooses the capturable link of the chain (packaging, testing, design, logistics) instead of promising wafer fabs. Meanwhile, Ecuador received the United Arab Emirates’ AI minister, Omar Sultan Al Olama, with an agenda that mixes infrastructure, energy, security and investment and with no agreements announced in advance: the Gulf has entered the region with the capacity to finance compute that no national budget covers, and with far less public discussion than Chinese or U.S. cooperation. Outside the region, two references worth following closely because they end up being copied: the transparency obligations of Article 50 of the European AI regulation took effect on August 3, and the first serious analysis, from Tech Policy Press on the same day, documents their gaps before anyone imports them: content provenance is only “encouraged,” there are exemptions for personal and real-time use, and labels are erased when content is re-encoded.
Launches
- Qwen3.8-Max — It processes documents of more than 200 pages and video of more than 100 hours. Already available via API on Alibaba Cloud Model Studio and QwenWork; the weights arrive on Hugging Face and ModelScope in mid-August. What is interesting is not the ranking but who can host it: a university or a ministry in the region, in its own server room. One question remains unanswered anywhere: how it performs in Spanish and Portuguese on genuinely long tasks, because its performance in our languages has not been reported.
- MiniMax H3, now with open weights — A 33-billion-parameter model that generates clips of 4 to 15 seconds with native stereo audio and accepts, in a single request, up to nine reference images, three video clips and three audio clips. According to Artificial Analysis, it ranked first in video editing and second in text-to-video: the first time an open model has topped a video ranking. The weights were published on August 3 under a community license, with the 2K module excluded and a practical ceiling of 768p when run locally. It matters because it is uncomfortable: synthetic video with voice, on one’s own hardware, with no mandatory watermark and no per-second cost, the same week Europe began requiring marking and with an election calendar looming in the region.
Threads we’re following
This adds to two stories we had been following that touch today. Yesterday we reported that Mexico now exports more AI servers than cars, capturing real money from the boom but in the most replaceable link of the chain, without connecting that manufacturing to its artificial intelligence agenda; Panama has just done exactly the opposite, with much less installed hardware and a single policy for both. And a few days ago we noted the European bet on sovereign compute: the European Commission opened the call for up to seven AI gigafactories, with 10 billion euros in public money leveraging 20 billion in private money, a minimum of 100,000 processors per facility, 18 member states buying jointly and a deadline of November 12. That (scale bought jointly among countries instead of isolated country-by-country competition) is still the instrument Latin America does not have.
If the model is no longer the bottleneck and the bottleneck is now where and under what rules the compute runs, does it make sense to keep writing artificial intelligence policy in science ministries rather than in finance and energy ministries?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.