Is this even stable?

[fryingpan]

I presume OP wants to stay discrete & bipolar..

If so, and taking into account his contenteness at amplification stage, a stand-alone (not directly coupled) BJT buffer could fulfill his requirements..

e.g. Emitter biased at about 5.5V with Emitter resistor of 3k3 could easily drive Q2 output of about 2.8V_RMS (4V amplitude) into a 10k load..

Yes. That's certainly an easier (and more economical) way. I was trying not to

[antonis]

That's certainly an easier (and more economical) way. I was trying not to

Then think of my last post "edit..

edit:
@fryinpan: I'd configure Q4 as current source (2 series diodes from Base to GND, delete R11, make R3 10k or so and R12 value set according to current taste..)

[fryingpan]

That's certainly an easier (and more economical) way. I was trying not to

Then think of my last post "edit..

edit:
@fryinpan: I'd configure Q4 as current source (2 series diodes from Base to GND, delete R11, make R3 10k or so and R12 value set according to current taste..)

Isn't this basically a different way to do what I did?

(And technically isn't it a current sink? Same difference, of course...).

[antonis]

Isn't this basically a different way to do what I did?

It's just a matter of taste..

(strictly speaking, one of the diodes (in close proximity to Q4) balances for V_BE thermal variations..)

[R.G.]

It's probably stable because the CFP pair doesn't have enough poles in its response to oscillate on its own, and adding the emitter follower adds another pole or two, but because it's gain-bandwidth is all used on bandwidth, not gain, the follower pole is way the heck up in frequency.

This also accounts for the peak in response at very high frequencies in the sim. I would expect the peak to be out near the GBW of the 2N3904, which OnSemi says is a minimum of 300MHz. Our old rue of thumb was that the phase change from a pole starts taking effect one decade before the actual nominal pole location, so it could be starting up at 30MHz, drizzling phase shift above that.

This is only rule of thumb speculation of course, but it seems to fit the facts. It might be interesting to remove the emitter follower and run a stability gain/phase sim on the first two stages to see what they do without the follower.