HT Sherpa

Field Guide · Week of 17 August 2026

This week’s open models — ranked, by country, and by whether they fit your computer.

Every Monday we re-read Hugging Face (downloads, likes, last update) and re-rank the top nine in each list. The writing is ours. The numbers are theirs. We do not list a 1.6-trillion-parameter model as a laptop.

Laptop and desktop · top 9

On your computer

These nine fit a machine you already own. Memory, not marketing, decides which one.

  1. United States · Google

    1

    Gemma 4 12B

    Google

    The honest everyday model for a 16 GB laptop. Clear writing, Apache license, no account required after download.

    Runs on your computer · 16 GB (4-bit) · Apache 2.0

  2. China · Alibaba

    2

    Qwen 3.6 27B

    Alibaba

    The best open writer you can still keep on one machine. Multilingual. Apache. The download numbers this week are enormous.

    Runs on your computer · 16–24 GB (4-bit ~17 GB) · Apache 2.0

  3. United States · OpenAI

    3

    gpt-oss 20B

    OpenAI

    OpenAI’s first open weights since GPT-2. Built to run in 16 GB. Apache license. Very widely downloaded.

    Runs on your computer · 16 GB (native 4-bit) · Apache 2.0

  4. China · Alibaba

    4

    Qwen 3.6 35B-A3B

    Alibaba

    A 35-billion-parameter model that only runs three billion at a time. Laptop-honest mixture-of-experts.

    Runs on your computer · 16 GB (4-bit) · Apache 2.0

  5. United States · Google

    5

    Gemma 4 26B A4B

    Google

    A larger brain that only wakes a few billion parameters at a time. The best local model on a 32 GB Mac this week.

    Runs on your computer · 24–32 GB (4-bit) · Apache 2.0

  6. United States · Microsoft

    6

    Phi-4

    Microsoft

    Small, sharp, MIT-licensed. Strong at exams and short reasoning. Less warm as a letter-writer.

    Runs on your computer · 12–16 GB (4-bit) · MIT

  7. Europe · France

    7

    Mistral Small 4

    Mistral

    Europe’s best laptop-class model this year. French lab, Apache license, mixture-of-experts that fits a strong machine.

    Runs on your computer · 24–32 GB (4-bit) · Apache 2.0

  8. China · DeepSeek

    8

    DeepSeek V4 Flash

    DeepSeek

    DeepSeek’s smaller V4. MIT license. Quantized, it can live on a strong laptop or a single workstation GPU.

    Runs on your computer · 32–64 GB (quantized) · MIT

  9. United States · Microsoft

    9

    Phi-4 Mini

    Microsoft

    The model we put on an 8 GB machine. It will not write a novel. It will draft and explain.

    Runs on your computer · 8 GB · MIT