Fix keep_alive support / persistent model sessions for Ollama provider to prevent model reloads between requests

[gregory-chatelier]

When using Ollama as a provider in opencode, the model appears to reload into memory for each request.

This significantly increases latency (especially for large models such as qwen3-coder:30b) and consumes unnecessary system resources.

Example configuration:

{
    "$schema": "https://opencode.ai/config.json",
    "provider": {
        "ollama": {
            "npm": "@ai-sdk/openai-compatible",
            "name": "Ollama (local)",
            "options": {
                "baseURL": "http://localhost:11434/v1"
            },
            "models": {
                "qwen3-coder:30b": {
                    "name": "qwen3-coder:30b",
                    "options": {
                        "keep_alive": -1
                    }
                }
            }
        }
    }
}

Behavior:

Each time opencode sends a new request to the Ollama endpoint, the model is unloaded and reloaded by the Ollama server.

This can add up to 60 seconds of overhead on every call for large models.

Direct requests to Ollama using its API or CLI (with keep_alive or OLLAMA_KEEP_ALIVE) can keep the model resident in memory — but this flag seems not to be propagated or respected through the opencode integration.

Expected behavior

When configured, models should remain loaded in memory for a configurable duration (e.g., keep_alive: 300) between requests.

The keep_alive value should be passed through to Ollama’s /v1/chat/completions or /v1/completions API payload, if supported.

Optionally, a global default could be set in the provider’s options.

The following didn't worked :

      "options": {
        "baseURL": "http://localhost:11434/v1",
        "keep_alive": 300
      }

[rekram1-node]

@gregory-chatelier try this:

    "ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ollama (local)",
      "options": {
        "baseURL": "http://localhost:11434/v1",
        "headers": {
          "keep_alive": -1
        }
      },
      "models": {
        "qwen3-coder:30b": {
          "name": "qwen3-coder:30b"
        }
      }
    },

[rekram1-node]

I tested, the headers do get sent if you do it this way^^

[gregory-chatelier]

Nice try, but without any success.

I already set an OLLAMA_KEEP_ALIVE env var to -1 and I'm pretty sure now the issue is not this keep alive directive. I guess there is an additional caller authentification that isn't consistent between requests. I use Ollama under Docker by the way.

[rekram1-node]

@gregory-chatelier I don't have any ollama experience but someone made a simple guide of using opencode + ollama + docker compose, have given people some suggestions from this: https://github.com/p-lemonish/ollama-x-opencode

Not sure if this is helpful at all but a lot of people have told me they found it helpful, it'd be worth adding certain parts to our docs but let me know if anythign there is helpful or not :)

[gregory-chatelier]

It didn't solve my issue but it's indeed a valuable resource I would recommend

[FlorinAndrei]

Any progress on this? It's quite annoying if you use an Ollama endpoint and you can't control the keepalive of the server.

It should be added to the provider options, not to the model options. In other words, the setting should apply to a whole Ollama endpoint.

[BrutchsamaJeanLouis]

major issue for my local LLMS

[FlorinAndrei]

It would be great if this was solved somehow. But, at this point, it looks like this is for Ollama to fix.

Ollama has two APIs: their native API, and the OpenAI API.

The Ollama native API has a way to set the TTL for a model, and it works. But OpenCode does not use the native API.

The OpenAI API also has a way to set the TTL, but Ollama ignores it, because it does not have a full implementation of the OpenAI API. This is the API that OpenCode uses.

There is a PR to fix the OpenAI API issues in Ollama:

ollama/ollama#11249

But it looks like they are ignoring it.

Please bug the Ollama team. They also have a Discord.

[github-actions]

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.