Models installed on Machine Vs llmfits Detection of what is installed

[GitLoader-Tfoundation]

I'm not sure I'm interpreting it correctly, but when I'm using LLMfit in a sub terminal window of my Hermes desktop app, I have a difficult time installing certain models due to the fact that after only installing three a majority of the listed compatible models show to be installed(which makes them unavailable to me for download). This is well over 30 and I only installed three. So I'm not exactly sure why llmfit detect that I have installed other models outside the three that I actually acquired. Can someone shed some light on this particular situation. Is it because the models that I have downloaded mirror a majority of the perimeter, tokens, modes, or like the use case?

[AlexsJones]

It very likely depends on how you installed them and their path, llmfit looks in common LMstudio, ollama, hf, llama-server locations.

How did you install them?

[GitLoader-Tfoundation]

Well, at first, I used the default directory to load and install the 3 models. This is when I notice the problem . I tried to refresh the list, and i tried terminating the TUI and reopened. I ended up changing that path to the Ollama models folder.

thank you for you time.

[GitLoader-Tfoundation]

the way i ended that might make it seem the problem is solve , it is not . the 'install' path is just different with the same problem . (there is only one other model that was acquired apart from LLMFIT that is sharing the same path as my ollama's model folder )

[AlexsJones]

(This might also be why you see more than you expect) but im happy to debug with you

[AlexsJones]

Thanks for the detail — you found a real bug, and it isn't about the path.

Ollama detection doesn't look at a models folder at all; it asks the Ollama daemon over HTTP (/api/tags), so moving the directory wouldn't change anything. It's also nothing to do with parameters, tokens, modes or use cases — it's name matching, and it was matching far too broadly.

What was happening: llmfit recorded each installed model under both its full tag and its bare family name (qwen3:8b → qwen3:8b + qwen3), and for any model without an explicit entry in its mapping table it also looked up that bare family name. So one install marked every model in the family as installed. Measured against the 9,250-model catalog:

you install	llmfit marked installed
qwen3:8b	238 models — including Qwen3-235B-A22B and hundreds of community fine-tunes
llama3.2:3b	8
gemma3:4b	1 (a 27B fine-tune)

Three installs → 247 models shown as installed. That matches your 30+ exactly: your view is filtered to models that fit your hardware, so you were seeing a slice of the same over-match. And because "installed" is what hides the download action, the practical effect is what you hit — models you never installed become undownloadable.

Fix: #863. Size is now the discriminator — a sized install (qwen3:8b) only ever matches that model, and a :latest install is resolved to the specific size via the parameter count Ollama already reports, instead of matching the whole family. After the fix those same three installs mark 4 models installed, all of them correct.

If you can share which three models you installed (ollama list), I'll double-check them against the new behaviour before it ships. One known gap that PR doesn't close: gemma3:4b doesn't detect google/gemma-3-4b-it yet — that's a missing mapping-table entry and a separate follow-up.

[GitLoader-Tfoundation]

The models that I have loaded from LLMFIT are as followed:

• NousResearch/Nouse-Hermes-2-Mixtral-8x7b-DP0
• NousResearch/Hermes-3-Llama-3.1-8B
• Qwen/qwen2.5-coder-7B-Instruct

All Marked as provided by Ollama

Thank you for the continued support !

[AlexsJones]

Sorry for the slow follow-up, @GitLoader-Tfoundation, and thanks for listing the three models.

I checked them against the fixed matching: all three have their own entry in the Ollama mapping table (nous-hermes2-mixtral:8x7b, hermes3:8b and qwen2.5-coder:7b), so each install should now mark only its own model as installed, not the rest of the family.

The fix shipped in v1.1.10 (#863), with a follow-up for size-less tags in v1.1.11 (#899). The current release is v1.1.16. After updating, you should see exactly those three as installed and everything else downloadable again.

If anything is still marked that you didn't install, paste the output of ollama list here and I'll trace it.