Skip to content
BENIBIX Learn

Gemma, Phi and gpt-oss

Level 2 · Names, sizes and licenses

The small open models that Google, Microsoft and OpenAI share, though their main models stay closed. What each runs on.

Course contentsLesson 7 of 13

After this lesson you can

  • Name the open models from three closed labs
  • Say what a small language model gives up
  • Pick a model that fits a laptop's memory

10 mincovers 4 glossary entries

Worth reading first: AI Model Names and Sizes

Which closed labs also share open models

Even labs whose main models are closed share some smaller open ones. Google shares Gemma, Microsoft shares Phi, and OpenAI shares gpt-oss. All three can be downloaded and run on your own machine. Phi-4, from December 2024, has 14 billion parameters. The small gpt-oss model runs in 16 gigabytes of memory.

Gemma is Google DeepMind's family of open models, a small sister to Gemini. The newest is Gemma 4. Its small sizes are built for phones, and its largest holds 31 billion parameters. It ships under the Apache 2.0 license, which allows business use. Older Gemma models came with Google's own terms.

Phi is Microsoft's family of small language models. Its makers trained it on carefully chosen text, much of it written to teach, so it reasons well for its size. Phi-4 comes under the MIT license. A model this size fits on a good laptop, but it knows fewer facts than a big one.

In August 2025 OpenAI shared gpt-oss, its first open-weight language model since GPT-2. It comes in two sizes. The big one, at about 117 billion parameters, fits on one graphics card with 80 gigabytes. The small one, at about 21 billion, runs in 16 gigabytes of memory. Both think before they answer, and both use Apache 2.0.

Why give a model away? An open model gets used, tested and built on by people a lab would never reach. It wins goodwill with builders. And a small open model does not compete with the lab's paid flagship, which stays closed and far stronger.

These are small language models, built to run where a big model cannot. A phone, a laptop or one server can hold them. They answer fast, cost nothing per use once downloaded, and keep your data on the device. In return, they make more mistakes on hard or rare questions.

People think a model from a closed lab must be closed. GPT, Gemini and Claude stay closed, but gpt-oss, Gemma and Phi do not. Check the exact model, not the lab. And do not look for gpt-oss inside ChatGPT. It is a download, run by you or by a host you pick.

One lab can sit on both sides. Google's Gemini is online only, and its Gemma is a file you keep.

Worked example

Emma studies with the internet off

Emma's school laptop has 8 gigabytes of memory. gpt-oss-20b needs 16, so it will not fit.

A small Gemma 4 model does fit. It answers her French questions on the train, with no internet at all.

The words for this lesson4

Check yourself

Can you download any model made by OpenAI?

Finished this lesson?

Marking it complete also marks its 4 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.