[tools] image failed: Unknown model: lmstudio/qwen/Qwen3.5-27b | Friends of the Crustacean 🦞🤝 | Page 1

silent prism Apr 23, 2026, 12:38 AM

#

Im running into an issue in the latest update, openclaw webchat cant read images:

[tools] read failed: ENOENT: no such file or directory, access '/root/.openclaw/workspace/avatars/img1.png' raw_params={"path":"/root/.openclaw/workspace/avatars/img1.png"}
17:23:22
[tools] image failed: Unknown model: lmstudio/qwen/Qwen3.5-27b raw_params={"image":"/root/.openclaw/workspace/imgage2.png","prompt":"Examine this image and describe what you see in detail - the subject, colors, style, mood, and any notable features."}

slim salmonBOT Apr 23, 2026, 12:38 AM

#

proper lanceBOT Apr 23, 2026, 12:38 AM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

slim salmonBOT Apr 23, 2026, 12:38 AM

#

silent prism Apr 23, 2026, 12:41 AM

#

Getting the following errors in openclaw logs:

[tools] read failed: ENOENT: no such file or directory, access '/root/.openclaw/workspace/avatars/img1.png' raw_params={"path":"/root/.openclaw/workspace/avatars/img1.png"}
17:23:22
[tools] image failed: Unknown model: lmstudio/qwen/Qwen3.5-27b raw_params={"image":"/root/.openclaw/workspace/imgage2.png","prompt":"Examine this image and describe what you see in detail - the subject, colors, style, mood, and any notable features."}

silent prism Apr 23, 2026, 4:09 AM

#

@tight bobcat if I move a photo into .openclaw/media/inbound it can read it in that path fine

tight bobcatBOT Apr 23, 2026, 4:12 AM

#

Let me check the docs/source paths first, then I’ll pin down whether this is a file-access thing or just the model selection biting you.

#

Yep, that’s a really useful clue.

If it works from .openclaw/media/inbound, then the file path/access part is basically fine. OpenClaw does stage inbound attachments into media/inbound/*, especially for sandboxed runs.

So the remaining problem is very likely the model ref, not the image path:

Unknown model: lmstudio/qwen/Qwen3.5-27b means OpenClaw could not resolve that exact model ID.
For LM Studio, the model ref has to match the exact key from http://localhost:1234/api/v1/models, including case.
If this is the image tool, the model also needs to be image-capable. If the primary model cannot accept images, OpenClaw uses agents.defaults.imageModel.

Can you paste these?

openclaw --version
openclaw models list
curl http://localhost:1234/api/v1/models

#

I’m especially looking for whether LM Studio reports the key as:

qwen/Qwen3.5-27b
or
qwen/qwen3.5-27b
or something else entirely

If openclaw models list does not show lmstudio/qwen/Qwen3.5-27b exactly, that explains the error.

Docs:
https://docs.openclaw.ai/providers/lmstudio
https://docs.openclaw.ai/concepts/models
https://docs.openclaw.ai/help/faq

silent prism Apr 23, 2026, 4:14 AM

#

OpenClaw 2026.4.15 (041266a)

Model Input Ctx Local Auth Tags
lmstudio/qwen/qwen3.5-27b text 128k yes yes default,image

"models": [
{
"type": "llm",
"publisher": "qwen",
"key": "qwen/qwen3.5-27b",
"display_name": "Qwen3.5 27B",
"architecture": "qwen35",
"quantization": {
"name": "Q4_K_M",
"bits_per_weight": 4
},
"size_bytes": 17468056915,
"params_string": "27B",
"loaded_instances": [
{
"id": "qwen/qwen3.5-27b",
"config": {
"context_length": 120000,
"eval_batch_size": 512,
"parallel": 4,
"flash_attention": true,
"offload_kv_cache_to_gpu": true
}
}
],
"max_context_length": 262144,
"format": "gguf",
"capabilities": {
"vision": true,
"trained_for_tool_use": true
},
"description": null,
"variants": [
"qwen/qwen3.5-27b@q4_k_m"
],
"selected_variant": "qwen/qwen3.5-27b@q4_k_m"
}

tight bobcatBOT Apr 23, 2026, 4:19 AM

#

I think this is two separate failures.

read ENOENT
That path just is not present in the active workspace/sandbox. For webchat/uploads, OpenClaw stages inbound media under media/inbound/<file>, so it working from .openclaw/media/inbound is expected, not a clue that general file permissions are broken.
image unknown model
This looks like the real bug. LM Studio reports the canonical key as:

qwen/qwen3.5-27b

but OpenClaw is trying:

lmstudio/qwen/Qwen3.5-27b

That casing mismatch is enough to cause Unknown model, and it smells like either stale image-model config or an LM Studio normalization regression in 2026.4.15.

Also, this line is suspicious:

#

lmstudio/qwen/qwen3.5-27b text ... default,image

LM Studio says that model has capabilities.vision: true, so I’d expect OpenClaw to treat it as image-capable, not plain text.

What I’d try first:

openclaw models add lmstudio qwen/qwen3.5-27b
openclaw models set-image lmstudio/qwen/qwen3.5-27b
openclaw gateway restart
openclaw models list

If it still tries lmstudio/qwen/Qwen3.5-27b after that, I’d call it a bug in 2026.4.15 and file it here with the exact outputs you pasted:
https://github.com/openclaw/openclaw/issues

Relevant docs:

#

https://docs.openclaw.ai/providers/lmstudio
https://docs.openclaw.ai/concepts/models

If you want, paste the output of openclaw models status after those commands and I’ll help sanity-check it.

silent prism Apr 23, 2026, 4:35 AM

#

Yup seems broken even after that change, can you give me a summary so I can open a bug?

tight bobcatBOT Apr 23, 2026, 4:35 AM

#

Yep, here’s a paste-ready summary.

Suggested title
LM Studio vision model fails in image tool on 2026.4.15, resolves qwen/qwen3.5-27b as Qwen3.5-27b

Suggested bug body

OpenClaw version: 2026.4.15 (041266a)

I’m seeing image-analysis failures with LM Studio in webchat.

Environment
- Provider: LM Studio
- LM Studio model key: `qwen/qwen3.5-27b`
- LM Studio `/api/v1/models` reports:
  - `"key": "qwen/qwen3.5-27b"`

#

  - `"capabilities": { "vision": true, "trained_for_tool_use": true }`

OpenClaw model list shows:
- `lmstudio/qwen/qwen3.5-27b                  text       128k     yes   yes   default,image`

Problem
When trying to analyze an image in webchat, I get:

1) read failure
[tools] read failed: ENOENT: no such file or directory, access '/root/.openclaw/workspace/avatars/img1.png' raw_params={"path":"/root/.openclaw/workspace/avatars/img1.png"}

2) image tool failure
[tools] image failed: Unknown model: lmstudio/qwen/Qwen3.5-27b raw_params={"image":"/root/.openclaw/workspace/imgage2.png","prompt":"Examine this image and describe what you see in detail - the subject, colors, style, mood, and any notable features."}

What looks wrong

#

- LM Studio reports the canonical model key as `qwen/qwen3.5-27b`
- OpenClaw is trying to resolve `lmstudio/qwen/Qwen3.5-27b`
- That suggests a model-id normalization/casing bug, possibly using the display name instead of the canonical key
- Also, OpenClaw shows the model as `Input: text` even though LM Studio reports `vision: true`

Expected
- OpenClaw should resolve the exact LM Studio key (`lmstudio/qwen/qwen3.5-27b`)
- The model should be treated as image-capable
- Webchat image analysis should work without manually moving files around

Actual
- `image` tool fails with `Unknown model: lmstudio/qwen/Qwen3.5-27b`
- `read` against arbitrary workspace paths fails, while staging under `.openclaw/media/inbound` behaves differently

Workarounds attempted

#[tools] image failed: Unknown model: lmstudio/qwen/Qwen3.5-27b