Four things worth checking before you buy anything: the drive you already own, the version you're already on, what your apps installed without asking, and whether the files you downloaded are what they claim to be.
Slotstream buys capacity from your SSD (capacity).
Slotstream is a Mac app that runs a 125-billion-parameter model, roughly 104GB of weights, on machines with as little as 16GB of memory. It keeps only the portion of the model currently in use resident and pages the rest off the internal drive as it goes, sizing the resident set to whatever hardware it finds. A 48GB Mac gets about 12 tokens per second out of a model that has no business fitting on it. Friday's edition noted the cheap path to local memory closing as DDR5 stock went pre-sold and Intel started hinting at a return to the memory business. The answer that arrived is software, not silicon. The trade is idle drive bandwidth for memory capacity, and sustained paging happens on an SSD that is soldered in.
Anthropic's 5.1 point releases move quietly (evaluation).
Anthropic shipped point releases of both frontier models, Fable 5.1 and Mythos 5.1, with a what's-new document and a system card. Point releases are where the interesting behavior changes hide. Tool-calling reliability and long agent runs shift in ways benchmark tables do not report, because benchmark tables measure the things that stayed the same. Most operators do not pick a model version. The tool picks it, and plenty of tools float to whatever is latest on the vendor's side. Whether the AI features in your CRM, your support desk, and your internal assistants pin a version or track latest has a documented answer in each case, and most teams have never gone and read it.
Codex ships an office suite inside (procurement).
Simon Willison went digging in the Codex desktop app's cache and found a complete copy of LibreOffice sitting inside it. OpenAI solved document conversion by bundling a headless office suite into the download rather than calling a service, which means docx and xlsx handling happens on the machine and the file never leaves the network. That is the useful half. The other half is that an AI desktop app now arrives carrying third-party software the IT policy never reviewed and the disk never budgeted for. Anyone approving these installs across a team has a short and answerable question in front of them: what else came along, and under what license.
Compressed model files aren't what they claim (evaluation).
Open-weight models get republished in compressed form by volunteers, and the filename states how aggressively each copy was squeezed. A teardown alleges that AtomicChat's widely used Qwen3.8 files do not carry the compression level their names claim, which explains both why they fit where other copies don't and why they run faster. That is the second such case in five days. The practical effect is narrow and annoying: two people comparing local setups by filename are not comparing the same thing, and a quality drop blamed on the model may belong to the file. Kaitchup published a measured sweep of Qwen3.8-27B across compression levels, which is the reference for how far down quality holds. The curve only helps if the file matches its label.
Nothing here requires a purchase. Start with the version check: open whatever AI tool touches your CRM, your support desk, or your internal assistants and find out whether it's pinned to a model version or tracking latest, because that answer decides whether this week's 5.1 point release already changed how the tool behaves without anyone telling you.