This paper describes a major project involving many groups and many scientists to work towards high resolution climate modelling. The title is perhaps a bit tentative with the reference to "kilometre scale modelling", as the reported modelling is still an order of magnitude away from the kilometre scale. 

The paper is very interesting from the organisational as well as the scientific point of view. It is a clear example of what can be achieved by bundling the expertise of many research groups and different types of expertise. Not only climate scientists and model developers are involved but also computer experts,  software engineers, and application oriented people. Also the organisation of such a large number of people is worth reporting. Goals are set for each stage and hackathons are organised to discuss the results. I turns out to be an effective hands-on approach. 

The science is very interesting with two rather contrasting atmospheric models and two ocean models. The key theme is the role of high resolution. The ICON atmospheric model has a minimalistic parametrisation and leaves cloud generation and convection to the dynamics of the model. The IFS model, has highly developed parametrisations and gradually reduces the convective parametrisation activity at very high resolution. The key question for the ocean model is: what benefit does explicit modelling of ocean eddies bring? 

The paper is very well written and carefully crafted. Given the strong scientific and organisational messages, the paper is well worth publishing. I recommend publication in its present form. The authors might want to make a few changes related to the comments below.  

A few minor comments: 

Line 252-254 

Interesting that it takes a comparison with another model to find a mistake in the use of c_p and c_v in the formulation of surface fluxes! 

Section 4.3.1

The discussion on soil moisture/precipitation feedback is very interesting and important. It is very undesirable to have positive feedbacks in a climate model that do not exist in nature. Such an erroneous feedback inevitably leads to the simulation of unrealistic extremes. A positive feedback is present in many large scale models with parametrised convection and the current paper (and previous papers) suggests that the sign of the feedback is reversed in convection resolving simulations. The question is whether it is a parametrisation issue or a resolution issue? The current paper is not conclusive. IFS has some parametrisation of convection also at high resolution, but its "effective resolution" may be less than in ICON. 

My question is about the water budget. In summer, the main moisture source over large continental areas (e.g USA east of the Rocky Mountains) comes from land evaporation. This suggests strong re-cycling. The soil shows spring/summer drying but there is also runoff. At high resolution with intermittent convection there is a negative feedback where dry areas are preferred to trigger convection (e.g. Taylor et al.). How this works out at larger scales is not clear. Somehow it must change the mean water budget. If convection at small scales has a preference for dry areas, it means that the dry soil can absorb more water. Does it imply that averaged over large areas there is less runoff? In other words, if less water is taken out of the system then the water is still available for the re-cycling process? Did you look at runoff in these simulations? 

Fig. 8

The legend is very confusing, and it takes some time to understand the logic: solid lines refer to the left scale and dashed lines to the right hand scale, colours refer to ICO-C4, IFS_F-C4 and CERES. CERES does not have data in the top figure whereas it is in the top legend. Perhaps it is better to replace the legends by titles in which the logic is described.  

Section 4.3.2

The discussion about stratocumulus is very interesting. ICON does much better than IFS in the area considered and with tuning of the global radiation budget. However, the lack of contrast between the cumulus and stratocumulus regime seems to be shared by the two models. Is it correct to say that ICON is better in the strato-cumulus regime, and worse in the cumulus regime? 

This is an interesting paper that gives a high level overview of the

next GEMS project.  It describes a large effort to debug, run and

tune km-scale coupled climate model.  It's one of the first efforts

of its kind at this resolution.  The work is broken into 4 cycles,

each analyzing progressively more mature versions of the coupled

model.  The final cycle, with the most mature version of the model was

then used to address, with partial success, four climate science

questions.  The authors looked at several well chosen aspects of the

simulations in Seciton 4.  The results in Section 4.2 are quite

interesting, and I have some minor comments on that below.  I

appreciated the nice discussion and encouraging results for

stratocumulus in Section 4.3.

I only have minor comments:

1.  Some of the issues related to the energy budget that were resolved

during the early cycles appears to be coding errors and bugs. Would it

have been more efficient to also have a low resolution version of the

coupled model to detect and fix these issues?  Although both these

models are not designed to run at typical CMIP style resolutions

(~100km), would such a configuration run and be sufficiently Earth-like

to be appropriate for addressing some software and numerical issues?

2.  The paper makes extensive use of the "km-scale" adjective, and

defines this as (for the atmosphere) "...using horizontal grid

spacings equal to or less than 10km globally".  This definition is a

little optimistic, and I think most atmosphere model developers would

not consider 10 km resolution "km-scale".

In the introduction, the authors then state that "Km-scale atmospheric

simulations, also referred to as storm-resolving simulations, resolve

deep convection, capturing mesoscale convective systems ..."  Some of

the references cited to support that Km-scale atmosphere (so here,

that implies a 10km atmosphere) are convection resolving are Peters

et al 2019, which is using a 2.5km and Becker et al 2021, which

concludes "Our results suggest that deep convection is not completely

resolved at a resolution of 9 or even 4 km"

Thus I found the introduction a little unclear in that it has claims

that probably apply to km-scale models running a 2.5km or finer

resolutions, and may not apply at 10 km resolutions.  In particular,

the implied claim that 10 km model can resolve deep convection and

mesoscale convective systems is probably not correct.  

3. Related to comment #2, in section 4.2, the authors claim, 

"The fact that ICON-C4 can represent the tropical rainbelt over land

and the Eastern Pacific indicates that a horizontal grid spacing of

the order of 10 km is sufficient to reproduce the structure of

precipitation in those regions, and that is possible with a minimum

set of parameterizations.", and, 

"The experiments conducted by Segura et al. (2024) and Takasuka et

al. (2024) indicate that fine-tuning subgrid-scale processes can

produce a correct representation of the tropical rainbelt without the

use of convective parameterization."

I thought this claim was not well supported.  The Segura and Takasuka

references were more nuanced in their claims, and were from much more

constrained prescribed-SST simulations.  My superficial takeaway after

reading section 4.2 was that one still needs convective

parameterization (like in IFS) at this resolution.  It would be good

if the authors can strengthen their arguments for this result,

especially for non-expert readers such as this reviewer.