For the example .log file, these are values of convergence: Maximum Force - 0.000006 - 0.000002 - NO.....RMS - Force - 0.000002 - 0.000001 - NO....Maximum Displacement - 0.000638 - 0.000006 - NO.......RMS - Displacement - 0.000129 - 0.000004 - NO.. As you can see they are rather close to converge..

You might want to take advantage of redundant coordinates, right now you are forcing Cartesian coordinates. This will significantly slow down your calculations, often adding more cycles. Ultrafine is already the default grid, no need to specify this, but it might be too small for this accuracy. If you want to converge very tight in your structure, your scf should be converged tighter. You are using diffuse functions, you will typically be more cycles then. Try a different functional, maybe the fault is in Minnesota...

@CodyAldaz I know, have found that already @Tyberius - i have visualized in Avogadro, it shows just the last step. Where else could I do that? So you suggest me to set MaxStep=N to some different value? (e.g. smaller like MaxStep=15?)

@Martin-マーチン I did forced cartesian for that way I was able to converge without Ultrafine/Verytight. Actually I'm performing benchmark studies so I have chosen several functionals and several basis functiosn. Thus, to have valuable results for research paper, I'd prefer to retain Minnessota functionals

If you need to force Cartesian coordinates, then your molecule might be prone to errors in the integration grid (underdescribed, to small). If the other methods converge, but M06 does not, then maybe this functional is incapable of describing your system, try one of the newer ones then. Set your cycles to 500, if it doesn't converge then, you have serious problems. Try using a new guess. Does your molecule have symmetry? If yes, use it. Occasionally unsymmetric attempts oscillate around the symmetrical and never reach the actual minimum. (Don't run compound jobs.)

@Martin-マーチン Would it be wise to check unpruned grid, Int=(Grid=199974)? Well, that's some kind of result the functional doesn't work though it should. M06 are widely used when describing compounds I am studying. Isn't 500 cycles too much, wouldn't it pinpoint the geometry is definitely wrong? How exactly new guess influences results? No, it has not.

Sometimes the scf is just not good enough to use the Berny algorithm, or projecting the gradient or whatnot. Then a fresh start might help circumventing this. Increasing the tightness of the scf might also help getting more reliable derivatives. You can try superfine grid, but it takes a lot longer. Without the molecule specification nobody will be able to reproduce your problem, but given that there is H, C, and O only, M06 should be appropriate. Many times 200 cycles is good, but it only tells you how well your first guess was. If you need more than 500 there's something serious.

Hello there : ) Stack has advised to move the discussion into the chat so I did

I have also attached .xyz geometry to the original post so one could check and repeat my calculations