Daily Briefing
4 August 2026 · 5 stories
🔥 Top story
01Alibaba opens the weights of a 2.4 trillion-parameter model: a freight train that can work on its own for 16 days
Picture an employee who never sleeps, never takes a day off, and at the end of sixteen days delivers software with 265 patches, 127 improvement requests and 151 logged bugs. That is what Alibaba just demonstrated with Qwen3.8-Max: an ensemble of experts that, on any single request, only uses 95 billion parameters out of 2.4 trillion. Available since 3 August on QwenCloud at a few dollars per million words. The free weights are promised on Hugging Face in the week of 10 August — the first Max-tier model ever delivered as open weights. It speaks the tool language of both OpenAI and Anthropic, which saves integrators a rewrite.
02MiniMax opens its first text-to-video AI in open weights, but not from your country
Want to create a fifteen-second video from a written instruction, with speech and music in the same pass? Chinese lab MiniMax makes the weights of its H3 model free to download. The big caveat: the licence explicitly forbids using it in the United States, the European Union, the United Kingdom and South Korea, and by any company above twenty million dollars of revenue without written authorisation. Read it as a warning shot: the definition of 'open' is getting more and more variable.
03A voice agent that keeps calling its tools while it speaks
Imagine a phone assistant that does more than just talk: while it answers you, it pulls your customer file, checks your stock, books an appointment. NVIDIA publishes under a research-only licence an 11-billion-parameter model that does listening, reasoning and speech synthesis in a single pass — where others stack three pieces of software. It answers in 450 milliseconds. For any company looking to industrialise outbound automated calls, this is the open-weight ancestor to test quickly.
04Half a billion dollars for a European AI chip made of light
British startup OLIX closes a 312 million dollar Series B to commercialise an AI chip that moves its calculations through light rather than electricity. The idea: shuffle data between processors through optical fibres, so the bill drops sharply. First delivery is set for late 2027. Reed Hastings (Netflix) and a British sovereign fund are on the cap table. For European customers looking to be slightly less dependent on Nvidia, it is a signal that the supply chain is starting to diversify.
05One cyber attack in two now uses AI, and the reaction window shrinks from days to hours
CrowdStrike's annual report puts a number on it: AI-assisted attacks jumped 89 % in a year. The window between a vulnerability disclosure and its exploitation has fallen to 24 to 48 hours; China-linked attackers exploit in less than 24 hours. Sobering figure: 131 dependencies of the same agent-AI tool were poisoned in a single day by a North Korean group. For companies deploying agents built on open-source libraries, the warning is clear: audit now, and rework the update flow from week-scale to day-scale.
📡 To watch
Texas freezes new AI data centres pending an electric grid audit
Roughly 1,800 projects are queued, representing five times the grid's peak load. Possible template for other states: Arizona, Virginia.
Qwen3.8-Max open-weights window: watch between 10 and 17 August
If Alibaba keeps the promise, it is the first time a Max-tier model lands as open weights. Check independent scores on Artificial Analysis as soon as it drops.
'Forbidden zone' Chinese licence cascade after MiniMax H3
If other labs adopt the pattern, 'open' Chinese weights turn into regionally restricted access. Bake that into your model strategy by client jurisdiction.
Immediate audit of npm dependencies for Mastra AI agents
131 packages compromised in a single day by North Korean group STARDUST CHOLLIMA (see CrowdStrike file). Check installed versions and pin your dependencies.
📊 Trend
'Open weights' no longer means 'accessible everywhere'. This week shows it twice: MiniMax explicitly excludes the United States, the European Union, the United Kingdom and South Korea from its H3 licence; Qwen3.8-Max promises open weights but only through its own channels. For European users, the reflex 'it's open, so I can use it' needs to become 'I check the licence'. Layer that with the Texas regulatory side, which adds a physical layer (grid electricity) to the moral debate about AI sovereignty. Keep both windows (legal and infrastructure) in mind for stack arbitrage over the next 6 to 12 months.