2026-08-04

1. MiniMax shipped the H3 weights — one checkpoint of three, and four jurisdictions excluded

Yesterday this brief carried MiniMax's unkept "coming days" promise as a Watch item. It was kept on 2026-08-03: MiniMax H3 weights are live at MiniMaxAI/MiniMax-H3, three days after launch. The License row has moved off unknown. What it moved to is more interesting than the fact that it moved.

What shipped is H3-Base — 33.1B parameters, 768p with native stereo audio, 42.5 GB across two checkpoints. H3-Context-IR and H3-Regenerate-2K were withheld; MiniMax says the first "relies on a multi-stage hosted workflow" and the second is not open-sourced "due to the complexity of the system". The consequence is flat: the 1440p output the model was launched and priced on cannot be reproduced from the open weights. MiniMax's own "Full 2K Workflow" pairs a local H3-Base with Open Platform API calls for the two missing modules.

The licence excludes territories, not uses. The MiniMax Community License is free below $20M revenue with "MiniMax H3" shown in the UI. Line 10 excludes the United States, European Union, United Kingdom and South Korea — MiniMax's stated reason being that those regions are "developing or enforcing AI-related regulations that may have specific implications for generative video models". Organisations there may apply for a formal licence.

Why it matters: the open-weights fight assumes a government deciding whether weights may be released. Here the lab excluded the regulating jurisdictions itself, pre-emptively, and the excluded set is exactly the regions writing generative-AI rules — weights downloadable worldwide and licensed nowhere in the West are neither side's model of the argument. Inkling (Apache 2.0) and Laguna S 2.1 (OpenMDW-1.1) restrict by neither territory nor component.

→ MiniMax H3, MiniMax, Open-Weights Policy Fight → Snapshot · Hugging Face · Licence Q&A · RuntimeWire

2. Qwen3.8-Max went generally available and disclosed the number it withheld for 15 days

Qwen 3.8 Max left its 2026-07-19 preview on 2026-08-03. Four spec rows that had read unknown since this page was created are now filled: 95 billion active parameters of 2.4 trillion total, a 1 million token context window, $2 / M input · $6 / M output, and availability through Alibaba Cloud Model Studio globally plus QwenWork, Alibaba's workplace agent platform.

The active count is the one this wiki flagged on 2026-07-20 as "not disclosed — a critical gap for assessing true inference cost". It took 15 days.

First independent measurement: Arena.AI places it second globally on multimodal tasks, behind only Claude Fable 5 — which agrees with Alibaba's July self-claim on one benchmark and one modality, and the July claim was not scoped to multimodal. As of that date it was listed on neither Artificial Analysis nor Hugging Face, so the Intelligence Index figure this wiki uses to compare frontier models still does not exist for it.

Open weights for Max and the new Qwen 3.8 27B are stated for the week of 2026-08-10 — the first date attached to a promise made without one on July 19.

Why it matters: both of the largest announced Chinese open releases of 2026 have now shipped an API first and the weights later or not yet. The policy argument treats "open-weight lab" as a standing property of a company; two months of releases say it is a promise with a date, and the date moves.

→ Qwen 3.8 Max, Qwen 3.8 27B (new), Alibaba / Qwen AI Lab, Open-Weights Policy Fight → Snapshot · Qwen on X · MarkTechPost · SCMP

3. A prototype virus that runs its inference on the machines it infects

Researchers at the University of Toronto, the Vector Institute, University of Cambridge and ServiceNow built a computer virus that uses AI models to compromise machines and then runs inference on the compromised machines' own GPUs, using that to work out how to spread further. The authors state that "self-sustaining AI-driven cyber-threats are no longer theoretical".

The reporter disputes his own headline. Jack Clark, who carried it in Import AI 467, is sceptical of "self-sustaining" on operational grounds: what counts as persistence that is not simply squatting on human infrastructure, when credentials expire, services rate-limit, machines reboot, and bills show up. Recorded unresolved — it is a laboratory prototype, and no in-the-wild instance of this class appears in any source this wiki holds.

Why it matters: every incident on AI-Enabled Cyberattacks before this one consumed inference from somebody's API, which is why account-level detection worked and why Anthropic could count 832 accounts and 482 techniques. A worm carrying its own weights on stolen compute has no account to disable — and it is the first capability argument here that turns on open weights being portable rather than inspectable.

→ AI-Enabled Cyberattacks, Open-Weights Policy Fight → Snapshot · Import AI 467