Are current on-device LLMs useless?

I found this model "lfm2.5-thinking" which is "designed for on-device deployment" and is really tiny: just 1.2b params (~700 MB)

My first and very short interaction was so terrible that I felt no need to continue, but not in the way you think:

Here is the extremely brief transcript:

    >>> hey
    Thinking...
    Okay, let's see what the user wrote here. They just said "heat". Hmm, maybe they want a response to that? ... First "heat" could be a standalone message...

You saw that? It misread my "hey" as "heat", and kept going like this throughout the entire thinking process.

I thought I know what to expect from tiny local models like this.

But I did not expect it to be this bad. How would this be useful for anything local if it cannot even read a three letter-word properly?

Is this just lfm2.5 being terrible, or is this level of error a common trait in this model size range. Curious to hear from other folks' experiences.