Gemma 2 2B Instruct

A GGUF repackaging of google/gemma-2-2b-it, Google’s 2B instruction-tuned Gemma 2 model, made runnable in Ollama.

This model is not my work. All credit for the weights, training, and architecture belongs to Google DeepMind. This page exists only to make their model one ollama run away.

Note on the name: this tag is gemma-2b, but the model inside is Gemma 2 2B Instruct — not the older Gemma 1 gemma-2b. Google’s naming (gemma-2b vs gemma-2-2b) is one hyphen apart and easy to misread. If you want Gemma 1, use Ollama’s official gemma:2b instead.

Credits

The weights are byte-identical to the upstream Q4_K_M quantization. Nothing was retrained, merged, fine-tuned, or altered.

Quick start

ollama run treyleo16/gemma-2b

API:

curl http://localhost:11434/api/chat -d '{
  "model": "treyleo16/gemma-2b",
  "messages": [{"role": "user", "content": "Explain gradient descent in one paragraph."}]
}'

Model details

Stop tokens <start_of_turn> and <end_of_turn> are set in the model’s params layer, so multi-turn chat terminates correctly without extra configuration.

What to expect

Gemma 2 2B punches above its size on instruction following and short-form reasoning, and it is small enough to run on CPU or modest GPUs. It is a direct-answer model — no thinking block, no tool calling. If you need reasoning traces or function calling at this size, look elsewhere.

The 8K context is the real constraint. It is comfortable for chat, summarization, and short document work, and it is not the model for long-document RAG.

License — read this before redistributing

This model is not Apache-2.0 or MIT. It is governed by Google’s Gemma Terms of Use, which carry obligations that MIT-style licenses do not:

• Use is subject to the Gemma Prohibited Use Policy.
• If you redistribute this model or a derivative, you must pass along the Gemma Terms of Use and the use restrictions to whoever receives it.
• You must include this notice: Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.

The full notice and attribution ship inside the model itself:

ollama show --license treyleo16/gemma-2b

Google gates the original repositories on Hugging Face behind license acknowledgment. Pulling from Ollama skips that step, so the responsibility for reading and following the terms is yours.