Implementing Ollama's /api/show

[dredwilliams]

I am trying to use olla between my model collection and LangFlow. When I configure the System Models to use Ollama and provide my olla URL, it accepts it as valid but shows no models.

In trying to diagnose the issue, I discovered that while LangFlow sends the /api/tags call to get the current set of models (which olla properly responds to), it then does an /api/show call to get model details. When it gets the "olla doesn't do /api/show" response, it assumes the model is invalid, and as a result has no models to show.

Is there a plan to implement this capability in olla?

Thanks!

[thushan]

Thanks for the great diagnosis @dredwilliams, that's exactly it. You've saved us a good chunk of digging :)

You're spot on: LangFlow probes /api/show after /api/tags to pull per-model details, and right now Olla deliberately returns a 501 there. That call was originally because reconciling modelfiles/parameters across multiple Ollama instances behind
the proxy gets messy, but for your use case that's overly cautious and it's clearly breaking a legit integration. So yes, we'll look at supporting it.

The plan we're leaning towards is to answer /api/show straight from Olla's unified model registry (the same way /api/tags works today) rather than fan out to a backend, which keeps it fast and avoids the multi-instance reconciliation problem.

To make sure we return enough for LangFlow to be happy, a few questions:

• What version of LangFlow are you running?
• What version of Olla? (just in case)
• What version of Ollama are your backends on?
• Could you grab the raw /api/show response from one of your Ollama nodes directly (bypassing Olla) for a model LangFlow accepts? Something like:

  curl http://<your-ollama>:11434/api/show -d '{"model":"<model-name>"}'
  

  That'll let us match the exact fields LangFlow validates so we don't ship a half-answer. Remove anything personal, this is the internet after all.

Will get a pre-release build to you to test once we've got something working. Thanks again for the detailed report!

[dredwilliams]

Thanks for picking this up!

I'm using LangFlow 1.9.6 and olla 0.0.27 with a custom mod to implement prioritized least-connections (see issue #84). Ollama is 0.30.5 all around (5 hosts running Linux and Windows).

The change for prioritized least-connections was simply to sort the model entries using the priority field -- just like is done for priority load balancing -- before the connections are evaluated. Did exactly what I wanted, and didn't require any structural changes elsewhere.

I'd include the output from /api/show ... but it is massive ... it includes the description, list of tensors, and requirements. I've excerpted key sections from the output for gemma4:31b ...

{
  "License": "Apache License ..."
  "parameters": "temperature                    1\ntop_k                          64\ntop_p                          0.95",
  "template": "{{ .Prompt }}",
  "details": {
    "parent_model": "",
    "format": "gguf",
    "family": "gemma4",
    "families": [
      "gemma4"
    ],
    "parameter_size": "31.3B",
    "quantization_level": "Q4_K_M"
  },
  "model_info": {
    "gemma4.attention.head_count": 32,
    "gemma4.attention.head_count_kv": null,
    "gemma4.attention.key_length": 512,
    "gemma4.attention.key_length_swa": 256,
    "gemma4.attention.layer_norm_rms_epsilon": 0.000001,
    "gemma4.attention.shared_kv_layers": 0,
    "gemma4.attention.sliding_window": 1024,
    "gemma4.attention.sliding_window_pattern": null,
    "gemma4.attention.value_length": 512,
    "gemma4.attention.value_length_swa": 256,
    "gemma4.block_count": 60,
    "gemma4.context_length": 262144,
    "gemma4.embedding_length": 5376,
    "gemma4.embedding_length_per_layer_input": 0,
    "gemma4.feed_forward_length": 21504,
    "gemma4.final_logit_softcapping": 30,
    "gemma4.rope.dimension_count": 512,
    "gemma4.rope.dimension_count_swa": 256,
    "gemma4.rope.freq_base": 1000000,
    "gemma4.rope.freq_base_swa": 10000,
    "gemma4.vision.attention.head_count": 16,
    "gemma4.vision.attention.layer_norm_epsilon": 0.000001,
    "gemma4.vision.block_count": 27,
    "gemma4.vision.embedding_length": 1152,
    "gemma4.vision.feed_forward_length": 4304,
    "gemma4.vision.num_channels": 3,
    "gemma4.vision.patch_size": 16,
    "gemma4.vision.projector.scale_factor": 3,
    "general.architecture": "gemma4",
    "general.file_type": 15,
    "general.parameter_count": 31273089132,
    "general.quantization_version": 2,
    "tokenizer.ggml.add_bos_token": false,
    "tokenizer.ggml.add_eos_token": false,
    "tokenizer.ggml.add_mask_token": false,
    "tokenizer.ggml.add_padding_token": false,
    "tokenizer.ggml.add_unknown_token": false,
    "tokenizer.ggml.bos_token_id": 2,
    "tokenizer.ggml.eos_token_id": 1,
    "tokenizer.ggml.eos_token_ids": null,
    "tokenizer.ggml.mask_token_id": 4,
    "tokenizer.ggml.merges": null,
    "tokenizer.ggml.model": "llama",
    "tokenizer.ggml.padding_token_id": 0,
    "tokenizer.ggml.pre": "gemma4",
    "tokenizer.ggml.scores": null,
    "tokenizer.ggml.token_type": null,
    "tokenizer.ggml.tokens": null,
    "tokenizer.ggml.unknown_token_id": 3
  },
  "tensors": [
    ...
  ],
  "capabilities": [
    "completion",
    "vision",
    "tools",
    "thinking"
  ],
  "modified_at": "2026-05-25T21:06:19.608869267Z",
  "requires": "0.20.0"
}

[thushan]

This is a quick WIP i've spent the evening playing with, as you've built yours off 0.0.27 you may be missing some other things too.

Do you want to see if this is workable?
https://github.com/thushan/olla/tree/feature/api-for-show

I still have to check for other issues but that may not be for a few days.

[dredwilliams]

Yes ... life happens ... I'll be able to test that tomorrow. Thanks!

[dredwilliams]

Sorry it took me so long to get back to this ... life gets complicated ...

The mod did work as expected -- LangFlow now recognizes all the connected models when accessing it through the /olla/ollama API endpoint. I also did a simple flow using a model through olla and it responded as expected.

Thanks!

Ed