Error messages

Error: 400 Bad Request: "embeddinggemma" does not support generate

Error: 400 Bad Request: "embeddinggemma" does not support generate

You pull a model, run it, and get a 400 back instead of a prompt:

$ ollama run embeddinggemma
Error: 400 Bad Request: "embeddinggemma" does not support generate

That's from issue #12182, filed on September 4, 2025. The model name in quotes changes (qwen3-embedding is another regular), and chat front ends that use the chat endpoint see does not support chat instead. As of September 26, 2026, 24 Ollama issues quote the first wording and 17 the second.

What the model can and can't do

An embedding model reads text and returns a long list of numbers (a vector) that captures its meaning, which is what search and retrieval systems store and compare. It never writes text. We read how Ollama 0.34.4 decides this: when a model file declares a pooling type (the step that squashes a text into one vector), Ollama marks it as an embedding model, and anything else is assumed to generate text.

The generate and chat endpoints both ask for the text-generation ability first. An embedding model hasn't got it, so the server stops with a 400 before loading anything. An Ollama contributor put it plainly in #12449: "qwen3-embedding is not a text generation model so it doesn't support generate."

Older Ollama refused these in ollama run

Until late 2025 there was no way to use an embedding model from ollama run at all, and that same contributor told the #12449 reporter "you can't ollama run it." Version 0.12.10 (November 5, 2025) changed that, and its release notes lead with "ollama run now works with embedding models". When someone hit the error again in December (#13557), a collaborator's whole answer was "Upgrade ollama", with console output from 0.13.5 printing a vector.

On a current version, ollama run spots the embedding model and prints its vector as a JSON array. It wants text to work on, though. Run it bare and 0.34.4 answers embedding models require input text with a usage hint.

ollama run embeddinggemma "Hello world"
echo "Hello world" | ollama run embeddinggemma

Calling it from code or a front end

Programs should use the embeddings endpoint. From Ollama's embeddings page:

curl -X POST http://localhost:11434/api/embed \
  -H "Content-Type: application/json" \
  -d '{
    "model": "embeddinggemma",
    "input": "The quick brown fox jumps over the lazy dog."
  }'

If a chat app shows does not support chat, check which model is selected in the chat picker. If the app also searches your documents, the embedding model belongs in that setting (it's the part that does the searching), and the chat picker needs a model that writes text.

If a model you expect to chat with gets this error (it has happened with starcoder and llava:7b), pull it again so you have the current version. That was the first question an Ollama contributor asked in the llava report.