Which AI Service?

[p3ck]

Hello,

I was wondering what others are using for AI service here? I tried using the Free Tier of Gemini and even basic queries hit the limit. I was considering setting up OOLAMA but I'd be curious which model gets the best results. I am reluctant to setup a credit card on these AI models as I've heard horror stories about them charging crazy $$

[iChenchus]

Hi p3ck,

I went through exactly the same thing — I started on Gemini's free tier and hit the rate limits constantly (lots of 503s too), so I switched to OpenRouter and haven't looked back.

A couple of things that might ease your credit-card worry: OpenRouter lets you use models tagged :free without spending anything. And even if you do add credit, it's prepaid — you top up a fixed amount (I put in $10), it's not a subscription, so there's no way for it to run up a surprise bill. The "crazy charges" horror stories are mostly about leaving an uncapped API key exposed; with a prepaid balance that simply can't happen.

I use the openrouter/free model setting, which auto-routes to whichever free model is available. For AudioMuse it works well for most tasks, and it costs nothing.

One honest caveat from my own experience: for sonic similarity requests (naming seed songs/artists), the quality comes from AudioMuse's sonic engine, not the AI — so a free model is perfectly fine there. Where a weaker model can struggle is with conceptual or ambiguous natural-language prompts, where the AI has to actually interpret what you mean. If you find a specific query giving poor results, that's when a better model might help — but for everyday use, free has been more than enough for me.

Self-hosting via Ollama is also a valid route if you'd rather keep everything local, but on modest hardware the models you can realistically run are weaker than what you get free through OpenRouter — so unless local-only is a hard requirement for you, OpenRouter is the easier win.

Hope that helps.

Edit: One more thing on OpenRouter: in my experience, having a small prepaid balance on the account raises the daily request allowance for the :free models — I topped up $10 and got a much higher daily limit. I couldn't find this clearly spelled out in their current docs, so treat it as "what worked for me" rather than an official figure — the exact thresholds may have changed. Worth checking OpenRouter's own pricing/limits page for what applies today.

[NeptuneHub]

Hi all,
thanks for asking this, so first able remember that LLM is used only in 2 functionality:

• Clustering => to give a name on the automated created playlist. Otherwise you can still use clustering but with tag based name;
• Instant Playlist => here the funcitonality is basically the LLM that "compose" more AudioMuse-AI functionality togheter. Basically here the AI have the goals of understanding the requested.

For Ollama local deployment, assuming you don't have expensive CPU like me (I have an rtx 5060) I focused on Qwen, what I saw is that qwen3.5:9b work good (and is the base for my test), but also qwen3.5:4b and qwen3.5:0.8b somehow do their job depending from the complexity of your request. The first one you also have on Openrouter under the name of qwen/qwen3.5-9b.

Other model could potentially work but off course I can't test every model and I can confirm. Could happen that some different model, or some updated version of this module, bring some new parameter that could bring AudioMuse-AI to not working. Now for example thinking model don't work good on this kind of task (the tools selection of the instant playlist) and hands up in an infinite loop. So say that could happen that if you try different model it doesnt work.

Could have value to say that AtlasCloud.ai, for which I want to hihglight that I don't get any reward and that I didn't had the time to test myself, just reached me in this days to highlight that their model are compatible under the standard OpenAI model. So if you're interested you can give a look. But let me repeat: I'm not sponsored by them; I don't get money by them; I still didn't had the chance to try their product. I only agree with them that I link them as compatible product and they link me with the idea to give them a try in the next days.