VLM 模型支持 Ollama 本地模型

[dreamer-zq]

考虑到成本问题，我想知道VLM模型所需要的最小Ollama 本地模型最小配置是多少？

[parnika0110]

If the goal is to keep the local setup as lightweight as possible, I don't think OpenViking defines a specific "minimum" Ollama model size. The important requirement is that the model used as the VLM supports vision, since OpenViking uses the VLM for semantic extraction and multimodal content understanding.

For a local starting point, I'd try a small vision model such as qwen3-vl:2b.

Run ollama pull qwen3-vl:2b

Then configure the VLM through Ollama/LiteLLM:

{ "vlm": { "provider": "litellm", "api_key": "ollama", "model": "ollama/qwen3-vl:2b", "api_base": "http://127.0.0.1:11434" } }

Then run openviking-server doctor to verify that the local configuration and model connectivity are working.

I'd treat qwen3-vl:2b as a practical low-resource starting point rather than an officially defined minimum. If you need better extraction quality, you can move to a larger vision model later. Make sure an embedding model is configured separately as well, since the VLM and embedding model serve different purposes.

[imronreviady]

目前 OpenViking 并没有定义一个官方的“最小 VLM 配置”，不过当前代码里其实已经有一个比较明确的 Ollama 本地基线可以参考。

现在的 setup wizard 测试使用的是：

{
  "vlm": {
    "provider": "litellm",
    "model": "ollama/qwen3.5:4b",
    "api_key": "no-key",
    "api_base": "http://localhost:11434",
    "extra_request_body": {
      "num_ctx": 16384,
      "think": false
    }
  }
}

这里 num_ctx: 16384 比模型大小本身更值得注意。OpenViking 的代码注释明确提到，Ollama 默认的 4096 context 会截断 OpenViking 大约 5k token 的 extraction prompt，所以即使模型能正常启动，如果 context 太小，实际抽取效果也可能有问题。

因此如果目标是低成本本地部署，我会把 qwen3.5:4b + 16k context 当作目前更贴近 OpenViking 当前实现的起点，而不是把某个 2B 模型视为官方最低要求。

至于最低 RAM / VRAM，OpenViking 本身没有规定固定数值，这部分还会取决于 Ollama 使用的模型和量化版本。