Smaller local models

  • Huawei releases tools for making local models smaller — The open-source toolkit gives people running models locally another way to fit them into available memory.
    • MACaQ reduces the precision used to store model weights, a process called quantization. Its documented paths cover llama.cpp, vLLM, PyTorch and image-generation tools.
    • The llama.cpp guide includes CPU and GPU workflows, Qwen examples and 2-, 3- and 4-bit formats. It asks for representative calibration text and separate evaluation text.
    • MACaQ-R refines an existing compressed model while keeping it compatible with standard llama.cpp. The guide requires the matching original model and a full-precision teacher file to create it.
    • Huawei reports that its Qwen3.6 example has the same 22.36 GB file size as its comparison and a lower divergence from the full-precision model. That measures similarity, not performance on every task.
    • Caveat: These quality comparisons come from the authors; check your own workload and the runtime requirements before replacing a working model. (HN 116 points · 26 comments · Huawei source and guide)

Meeting notes on your own machine

  • Yogurt records and summarizes meetings with local models — The free, open-source app offers a way to keep meeting audio and transcripts on your computer.
    • The app captures microphone and system audio, transcribes with Whisper and summarizes with a local model. Its README says recordings, notes and transcripts are stored locally; models and update checks require internet access.
    • It lists macOS, Windows and Linux support and at least 8 GB of memory. The initial notes model is about 5 GB, with a larger model potentially 14 GB or more.
    • Optional access for assistant apps is off by default. Once enabled, apps running under the same account can read meetings, and an assistant may send what it reads to its own online model.
    • Caveat: The app records without a meeting bot, so tell other participants and follow applicable recording-consent rules. (HN 252 points · 96 comments · Yogurt README)

Limits on automated travel booking

  • United moves to block unauthorized AI assistants from booking flights — A travel assistant may run into the airline’s own rules even when a customer wants it to book.
    • United says it is restricting unauthorized third-party personal AI agents, meaning models carrying out multi-step tasks with tools, directly on its platforms and indirectly through travel agencies and aggregators.
    • The airline cites risks from sharing credentials and unintended purchases, account changes or use of miles. It also raises concerns about check-in requirements for travelers.
    • CBS quotes an aviation analyst who sees a commercial incentive to retain control of bookings as well. The analyst says ordinary aggregators will still be able to offer United fares.
    • Caveat: The reported policy targets unauthorized personal assistants; it does not establish that every automated travel service is prohibited. (HN 569 points · 581 comments · CBS News)
6 items considered but skipped