Open models worth using today

What you'll see in this lesson

1. Open model: you download the whole model

   An open model publishes its weights: you download them, run it on your machine or server, and your data goes through no one.

2. What came out in 2026, month by month

   Every big open lab has released a new version this year; this is what's current as of October 2026.

3. For a laptop or a 16 GB graphics card

   Gemma 4 and gpt-oss-20b are the open models built for regular machines, and both have the Apache 2.0 license.

4. Qwen3.8-27B: the sweet spot for 24 GB

   A dense model with 27.8 billion parameters, vision, a 262,000-token context and the Apache 2.0 license.

5. Mid-size MoEs: for a big Mac or a server

   gpt-oss-120b, DeepSeek V4-Flash and GLM-5.3-Flash are a leap in quality, but need from 65 to over 300 GB.

6. The giants: better via API or on rented servers

   Kimi K3, Qwen3.8 Max, DeepSeek V4-Pro and GLM-5.3 are the most powerful open models, but they won't fit at home.

7. What each one is good for

   There's no winner for everything: choose based on the task, the type of data and the memory you have.

8. Licenses: what you can do in client work

   Apache 2.0 and MIT let you use them in commercial work without asking permission; custom licenses add conditions.

9. Summary

   Choose the model by the memory you have, the task you need and what its license allows.