Large cardinals without the ambient set theory?  In an attempt to understand a bit better large cardinals, I have been thinking along the following lines, which could be summarized under the slogan

Talk about  cardinals without the
  (ambient) set theory

the class ON is first-order axiomatizable, and thus it looks like I can carve out of ON the subclass  CARD (for instance, one could add to the theory an equivalence relation,  $\alpha \equiv \beta$ formalizing  equinomerosity, and then define a cardinal in the usual way as the min ordinal in the equivalence class). 
Once I have my definable predicate $CARD(\alpha)$, I can proceed to introduce cardinal arithmetics. For instance, I can define successor as the minimal cardinal greater than the given cardinal.  
Obviously, I need to make some assumptions as to the basic cardinal arithmetics, so that it looks like the standard one in $ZFC$ + (possibly) generalized continuum hypothesis. 
Now, assuming one has done all of the above, it appears  that the "small" large cardinals, such as weak inaccessible, Mahlo, etc are definable in this theory (even in standard presentations, such as Drake,  their definition is arithmetic ). 
But what about the others, the heavy-weight ones? Do I necessarily have to  resort to the ambient set theory ( stationary points, elementary embeddings, etc ) to talk about very large cardinals, or there  is always a direct (algebraic/arithmetical/topological) way to provide their definition? 
Prima facie, it looks like the answer is no, but maybe there is a clever path to answer in the affirmative. Or perhaps, there is some kind of intrinsic boundary, beyond which you need to think of cardinals within the context of set theory
Any thought, refs, or known fact? 
 A: Most of the large cardinals have a variety of equivalent
formulations in ZFC, and some of them are characterized by long
lists of diverse equivalent properties. For a few examples, take a
look at the Cantor's Attic entries for weakly compact,
strongly compact and weakly measurable cardinals.
There is a phenomenon, however, that in weaker set theories such
as ZF some of these characterizations are no longer equivalent.
For example, one cannot expect to prove the embedding
characterization of measurable cardinals from the ultrafilter
characterization in ZF, since one needs the axiom of choice in
order to establish the Los theorem on ultrapowers. A similar situation arises with most all of the larger large cardinals. There has been some work understanding the various large cardinals in ZF worlds, for example, under the axiom of determinacy, which implies that many successor cardinals including $\omega_1$ are measurable according to the ultrafilter definition. 
Secondly, most of the largest large cardinals are defined by second order
properties, such as the existence of proper class objects, often
embeddings of a particular kind. These definitions cannot be made
in ZFC, which has no capacity for second-order quantifiers, but
rather are usually made in Gödel-Bernays set theory or
Kelly-Morse set theory. It turns out, however, that these
second-order definitions in every case (except Reinhardt
cardinals) have a first-order equivalent. For example, a cardinal $\kappa$ is measurable if and only if
(second-order) it is the critical point of an elementary embedding
$j:V\to M$, if and only if (first-order) there is a
$\kappa$-complete ultrafilter on $\kappa$. These equivalences are
not provable in ZFC, and not even statable in ZFC, but are proved
in GBC.
Meanwhile, the theory $\text{ZFC}^-$, meaning ZFC without the
power set axiom (see my paper, What is the theory of ZFC without
power set axiom? for some subtleties about how to axiomatize
this theory properly) is sufficient to formalize the basic
properties of those large cardinals with a $\Sigma_2$ definition,
such as inaccessible, Mahlo, weakly compact, measurable,
superstrong, Woodin and huge cardinals, whose existence is
absolute between $H_\delta$ and $V$. Meanwhile, notions such as
tall, strong and supercompact cardinals involve another quantifier
through all the ordinals, which makes them not usually absolute
between $H_\delta$ and $V$.
Although you seem to object to the embedding characterization of
large cardinals, I would say that it is the embedding
characterizations that have proved the most fruitful in their use
and analysis. It is surely the embedding outlook that unifies an
enormous part of the theory.
A: But anyway, my point is this: just look at ON and totally forget that is the spine of V, and see it as a system of ordered numbers. Now axiomatize what you see, much in the same way as you axiomatize its initial segment N. – Mirco Mannucci Jul 8 at 22:21
for instance, one could add to the theory an equivalence relation, α≡β formalizing equinomerosity, and
talk about cardinals without the (ambient) set theory
Well, this construction [Gavrilovich and Hasson’s Exercices de Style, a homotopy theory of set theory] attempts 
to talk about cardinal invariants and 
use as little set theory as possible: instead it uses axioms of a model category to do that. 
and indeed, as you suggest, there equinomerosity (up to some fixed $\kappa$) is introduced, under the name of cofibrant. The construction does not work with the skeleton, though; but 
perhaps it would be closer to using the skeleton if you modify  the defitinions and replace everywhere 'inclusion' aka 'subset' by 'injective map'; then you'd lose limits in your model category. 
How powerful the language is, is unclear. But you can indeed say $\kappa$ is measurable: 
that's when the corresponding homotopy caetgory is not dense as a partial order. 
