Ollama API: '"all-minilm" does not support generate' (embedding model)

{"error":"\"all-minilm\" does not support generate"} — the Ollama API answers this with HTTP 400 when you send a prompt to /api/generate (or /api/chat) for a model that can only make embeddings. The model is fine; you called the wrong endpoint. Send the text to /api/embed instead and you get a vector back.

Embedding models such as all-minilm and nomic-embed-text turn text into a list of numbers; they cannot write an answer. Use /api/embed with an input field for them, and keep /api/generate and /api/chat for chat models such as qwen2.5 or llama3.1.

The error

PS> curl.exe -s -i http://127.0.0.1:15190/api/generate -d '@gen.json'   (gen.json: {"model":"all-minilm","prompt":"Why is the sky blue?","stream":false})
HTTP/1.1 400 Bad Request
Content-Type: application/json; charset=utf-8
Date: Fri, 06 Nov 2026 07:15:06 GMT
Content-Length: 52

{"error":"\"all-minilm\" does not support generate"}

The chat endpoint refuses the same model with the same shape of message:

{"error":"\"all-minilm\" does not support chat"}

Why it happens

ollama show prints a Capabilities list for every model. The chat model in the test, qwen2.5:0.5b, lists completion and tools; all-minilm lists only embedding, no completion. Sent to /api/generate or /api/chat, that model was refused with HTTP 400 and its name in the message.

PS> ollama show all-minilm
  Model
    architecture        bert
    parameters          23M
    context length      512
    embedding length    384
    quantization        F16

  Capabilities
    embedding
...

The usual way to meet it is a RAG sample that uses one model name for both jobs, or a configuration where the chat model and the embedding model were swapped. On the command line Ollama 0.40.14 is more forgiving than the API: ollama run all-minilm "Why is the sky blue?" did not fail in the test, it printed the 384 numbers of the embedding and exited with code 0. So a script that works in the terminal can still fail the moment the same model name goes into an HTTP call.

The fix

PS> curl.exe -s http://127.0.0.1:15190/api/embed -d '@emb.json'   (emb.json: {"model":"all-minilm","input":"Why is the sky blue?"})
{"model":"all-minilm","embeddings":[[0.010121205,-0.0017384731,0.050096255,0.046960793,0.054998174,0.008553368,0.105450615,-0.025881601,0.12957646,0.031904623,- ... (cut) ... 354621]],"total_duration":9713000,"load_duration":3901200,"prompt_eval_count":8}

Two things changed: the endpoint is /api/embed, and the text goes in input rather than prompt. input also takes an array of strings: sent two questions in one array, the server returned two vectors. If you really wanted an answer in words, keep /api/generate and change the model to a chat model; in an application, keep two settings, one for the chat model and one for the embedding model, so they cannot be mixed up.

How it was reproduced

Ollama 0.40.14 on Windows 11, with a separate server started by ollama serve on 127.0.0.1:15190 (set through OLLAMA_HOST) and its own models folder, so the test server and port are the test's own. all-minilm (45 MB, embedding only) and qwen2.5:0.5b were pulled into it. The request bodies were saved to gen.json, chat.json and emb.json and sent with curl.exe; the generate and chat calls returned 400, the embed call returned the vector shown above.

Frequently asked

What does 'does not support generate' mean in Ollama?

The model you named has no completion capability. It is an embedding model, so /api/generate and /api/chat refuse it with HTTP 400. Send the text to /api/embed instead, or pick a chat model.

How do I get embeddings from Ollama?

POST to /api/embed with a JSON body that has model and input, for example {"model":"all-minilm","input":"some text"}. The answer has an embeddings array; in the test, an input array of two strings returned two vectors.

How do I check whether an Ollama model can chat or only embed?

Run ollama show with the model name and read the Capabilities section. A chat model lists completion; an embedding model such as all-minilm lists only embedding.

More decoded errors in the Fixes category. Every Ollama endpoint, with curl and .NET, is walked through in the Ollama API post.